CN106846361B

CN106846361B - Target tracking method and device based on intuitive fuzzy random forest

Info

Publication number: CN106846361B
Application number: CN201611170877.0A
Authority: CN
Inventors: 李良群; 李俊; 谢维信; 刘宗香
Original assignee: Shenzhen University
Current assignee: Kunshan Ruixiang Xuntong Communication Technology Co Ltd
Priority date: 2016-12-16
Filing date: 2016-12-16
Publication date: 2019-12-20
Anticipated expiration: 2036-12-16
Also published as: CN106846361A

Abstract

The invention discloses a target tracking method based on an intuitive fuzzy random forest, which comprises the following steps: performing motion detection on the current video frame, wherein a possible motion object obtained by detection is used as an observation result; correlating the observation result with a prediction result of a target, wherein the target comprises a reliable target and a temporary target; performing track management on the observation result and the prediction result which are not associated, wherein the track management comprises the steps of performing online tracking on the prediction result of the reliable target which is not associated to obtain a candidate result, and matching the candidate result by utilizing an intuitionistic fuzzy random forest of the reliable target which is not associated; and acquiring the track of the target of the current frame by using the correlation result and the matching result, predicting by using the track of the target of the current frame, and updating the intuitive fuzzy random forest for the reliable target which is successfully correlated or matched. The invention also discloses a target tracking device based on the intuitive fuzzy random forest. Through the mode, the target tracking performance can be improved under the condition of missing detection.

Description

Target tracking method and device based on intuitive fuzzy random forest

Technical Field

The invention relates to the field of target tracking, in particular to a target tracking method and device based on an intuitive fuzzy random forest.

Background

On-line target tracking is a hot research topic in computer vision, has important significance for high-level visual research such as action recognition, behavior analysis and scene understanding, and has wide application prospects in the fields of video monitoring, intelligent robots, human-computer interaction and the like.

In a complex scene, due to the influence of factors such as the deformation of the targets, mutual shielding among the targets, shielding of static background objects on the targets and the like, missing detection is difficult to avoid. At this time, the missed detection target cannot find the detected observation object associated therewith, and effective information cannot be found for the trajectory update of these missed detection targets through data association, and the trajectory accuracy is reduced.

Disclosure of Invention

The invention mainly solves the technical problem of providing a target tracking method and device based on intuitive fuzzy random forest, which can solve the problem of low track precision of missed targets in the prior art.

In order to solve the technical problems, the invention adopts a technical scheme that: the target tracking method based on the intuitive fuzzy random forest is provided, and comprises the following steps: performing motion detection on the current video frame, wherein a possible motion object obtained by detection is used as an observation result; correlating the observation result with a prediction result of the target, wherein the prediction result is obtained by predicting at least the track of the target of the previous video frame, and the target comprises a reliable target and a temporary target; performing track management on the observation result and the prediction result which are not associated, wherein the track management comprises the steps of performing online tracking on the prediction result of the reliable target which is not associated to obtain a candidate result, and matching the candidate result by utilizing an intuitionistic fuzzy random forest of the reliable target which is not associated; obtaining the track of the target of the current frame by using the correlation result and the matching result, wherein the track is obtained by filtering and updating the prediction result of the successfully matched reliable target by using the successfully matched candidate result; and predicting by utilizing the track of the target of the current frame, and updating the intuitive fuzzy random forest for the successfully associated or matched reliable target.

In order to solve the technical problem, the invention adopts another technical scheme that: the target tracking device based on the intuitive fuzzy random forest comprises: the processor is connected with the camera; the processor is used for carrying out motion detection on the current video frame acquired from the camera, and a possible motion object obtained by detection is used as an observation result; correlating the observation result with a prediction result of the target, wherein the prediction result is obtained by predicting at least the track of the target of the previous video frame, and the target comprises a reliable target and a temporary target; performing track management on the observation result and the prediction result which are not associated, wherein the track management comprises the steps of performing online tracking on the prediction result of the reliable target which is not associated to obtain a candidate result, and matching the candidate result by utilizing an intuitionistic fuzzy random forest of the reliable target which is not associated; obtaining the track of the target of the current frame by using the correlation result and the matching result, wherein the track is obtained by filtering and updating the prediction result of the successfully matched reliable target by using the successfully matched candidate result; and predicting by utilizing the track of the target of the current frame, and updating the intuitive fuzzy random forest for the successfully associated or matched reliable target.

The invention has the beneficial effects that: the method comprises the steps of carrying out online tracking on a prediction result of a reliable target which is not associated to obtain a candidate result, matching the candidate result by utilizing an intuitive fuzzy random forest of the reliable target which is not associated to obtain a track, and if the matching is successful, carrying out filtering updating on the prediction result of the reliable target by utilizing the candidate result which is successfully matched to obtain the track of the reliable target, so that the intuitive fuzzy random forest can be used for finding out the matched candidate result which can be used for filtering updating of the track of the reliable target under the condition that the target cannot find out an associated observation object due to the occurrence of missing detection, thereby improving the precision of the target track and improving the performance of target tracking.

Drawings

FIG. 1 is a flow chart of a first embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention;

FIG. 2 is a schematic diagram of a hard decision function and a fuzzy decision function of a branch node in an example of a second embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention;

FIG. 3 is a schematic diagram of a fuzzy decision function and an intuitive fuzzy decision function of a branch node in an example of a second embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention;

FIG. 4 is a flowchart of a third embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention;

FIG. 5 is a flowchart of a fourth embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention;

FIG. 6 is a flow chart of feature selection criterion training in a fourth embodiment of the target tracking method based on intuitive fuzzy random forest according to the present invention;

FIG. 7 is a flowchart of a fifth embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention;

FIG. 8 is a flowchart of a sixth embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention;

FIG. 9 is a flowchart of a seventh embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention;

FIG. 10 is a schematic structural diagram of a first embodiment of the target tracking device based on an intuitive fuzzy random forest according to the present invention;

fig. 11 is a schematic structural diagram of a second embodiment of the target tracking device based on the intuitive fuzzy random forest according to the present invention.

Detailed Description

As shown in fig. 1, a first embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention includes:

s1: and carrying out motion detection on the current video frame.

And performing motion detection on the current video frame by using motion detection algorithms such as a frame difference method, an optical flow method, a background subtraction method and the like to find out pixels belonging to a motion foreground, and finally obtaining a possible motion object in the current video frame as an observation object by using median filtering and simple morphological processing. An observed object is an image block in the current video frame, and in general, the observed object has a rectangular shape.

S2: and correlating the observed result with the predicted result of the target.

The targets include reliable targets for stable tracking and temporary targets for unstable tracking. The target status in this step, i.e. whether each target is marked as a reliable target or a temporary target, is determined by the trajectory management of the previous video frame. The temporary targets include new targets created for observations not associated with a previous video frame and not candidates for a successful match, and targets that have not been deleted for which the number of consecutive successful associations is less than or equal to a first threshold number of frames. The reliable target comprises a target which has a number of frames successfully associated with the target and is not deleted, wherein the number of frames successfully associated with the target is greater than a first frame number threshold. The prediction result of the object is obtained by performing prediction using at least the trajectory of the object of the previous video frame.

S3: and carrying out trajectory management on the observation result and the prediction result which are not associated, wherein the trajectory management comprises the steps of carrying out online tracking on the prediction result of the reliable target which is not associated to obtain a candidate result, and matching the candidate result by utilizing an intuitive fuzzy random forest of the reliable target which is not associated.

Specifically, a plurality of image blocks are selected as candidates within a predetermined range around the predicted result position of the reliable target, the size of the image block generally matches the size of the predicted result, and the size of the predetermined range and the number of candidates are generally determined by empirical values. The candidates may include observations within a specified range that are not associated. Adjacent candidates may not overlap each other, or may partially overlap each other. An intuitive fuzzy random forest of reliable targets that are not associated is used as a classifier whose classification results are of both reliable and unreliable target types. And calculating the candidate result as the intuitive fuzzy membership of the test sample to the reliable target class. If the intuitive fuzzy membership is greater than the first threshold eta₁And the similarity measure of the appearance characteristics of the candidate result and the prediction result of the reliable target is greater than a second threshold eta₂If the matching is successful, 0.5<η₁<1 and 0.5<η₂<1。

When the number of the candidate results is more than one, after calculating the intuitive fuzzy membership degree of each candidate result as the test sample to the reliable target class, whether the intuitive fuzzy membership degree of each candidate result is more than a first threshold eta can be respectively judged₁Whether the appearance feature similarity measure with the prediction result is greater than a second threshold η₂If at least two candidate results meet the two conditions, selecting one of the at least two candidate results with the largest intuitive fuzzy membership (selecting one with the largest appearance characteristic similarity metric if the intuitive fuzzy membership is the same) as a candidate result with successful matching for subsequent updating of the target state and the intuitive fuzzy random forest; of course, it is also possible to select a candidate target with the largest intuitive fuzzy membership from the candidate results, and then determine whether the intuitive fuzzy membership of the selected candidate result is greater than the first threshold η₁The similarity measure of the appearance features with the predicted result isWhether or not it is greater than the second threshold value eta₂And if the two conditions are met, the matching is successful.

In addition, the target state is updated according to the association result and the matching result, including the establishment, deletion and state modification of the target. The method specifically comprises the following steps: establishing a new temporary target for observations that are not associated and are not candidates for successful matching; the number of the successfully continuous correlated frames is larger than the threshold value lambda of the first frame number₁Becomes a reliable target; the number of unsuccessfully deleted continuous association frames is greater than the threshold lambda of the second frame number₂A temporary target of (2); the number of unsuccessfully deleted continuous association frames is greater than the threshold lambda of the third frame number₃And the matching result is a reliable target with matching failure, and the matching result is the matching failure means that the intuitionistic fuzzy membership degree of the test sample belonging to the reliable target class is less than or equal to a sixth threshold eta by using the intuitionistic fuzzy random forest calculation candidate result as the test sample₆And satisfy 0<η₆≤η₁. Wherein λ₁Is a positive integer greater than 1, λ₂And λ₃Are all positive integers and satisfy lambda₃≥λ₂≥1。

S4: and acquiring the track of the target of the current frame by using the correlation result and the matching result, predicting by using the track of the target of the current frame, and updating the intuitive fuzzy random forest for the reliable target which is successfully correlated or matched.

And filtering and updating the predicted result of the successfully matched reliable target by using the successfully matched candidate result to acquire the track. And filtering and updating the prediction result of the target which is successfully associated by using the associated observation result to obtain a track, taking the corresponding observation result as the track for a new temporary target, and taking the prediction result of the temporary target which is unsuccessfully associated and not deleted and the reliable target which is unsuccessfully associated and unsuccessfully matched and not deleted as the track.

And then, predicting by using the track of the target of the current frame, wherein the obtained result can be used as the prediction result of the target for target tracking of the next frame. In an embodiment of the present invention, a kalman filter is used to predict the trajectory of the target of the current frame to obtain the prediction result of the target of the next frame, and the kalman filter may also be used to filter the prediction result and the corresponding observation/candidate result to obtain the trajectory of the target.

And updating the intuitive fuzzy random forest for the reliable target by using the target image block corresponding to the successfully associated or matched reliable target. The target image block may not include track information of the target, such as a successfully associated observed object or a successfully matched candidate, and the execution sequence of the step of updating the intuitive fuzzy random forest and the step of acquiring and predicting the track of the target is not limited. The target image block may also include track information of the target, for example, an image block of a position where the track of the reliable target is located, and the step of updating the intuitive fuzzy random forest should be performed after the step of acquiring the track of the target.

Through the implementation of the embodiment, the prediction result of the reliable target which is not associated is tracked on line to obtain the candidate result, the intuitive fuzzy random forest of the reliable target which is not associated is used for matching the candidate result, if the matching is successful, the candidate result which is successfully matched is used for filtering and updating the prediction result of the reliable target to obtain the track of the reliable target, so that the intuitive fuzzy random forest can be used for finding out the matched candidate result which can be used for filtering and updating the track of the reliable target under the condition that the target cannot find the associated observation object due to the occurrence of missing detection, thereby improving the precision of the target track and improving the performance of target tracking.

The second embodiment of the target tracking method based on the intuitive fuzzy random forest is that on the basis of the first embodiment of the target tracking method based on the intuitive fuzzy random forest, the intuitive fuzzy membership degree P (c is m | w) of a candidate result serving as a test sample and belonging to a reliable target class is as follows:

wherein c is a class label of the test sample, m is a reliable target class, w is the test sample, and T is an intuitionistic fuzzy decision tree in an intuitionistic fuzzy random forestNumber, M, is a set of classes, φ, of training samples that generate an intuitive fuzzy random forest_tAnd (c ═ m | w) is the intuitive fuzzy membership degree of the test sample w to the reliable target class m calculated by using the t-th intuitive fuzzy decision tree.

The intuitive fuzzy decision tree performs intuitive fuzzification on branch node output judgment, so that the same sample can pass through an output left branch and an output right branch of a branch node with different intuitive fuzzy membership degrees and finally reaches a plurality of leaf nodes. Therefore, the classification result of the intuitive fuzzy decision tree needs to comprehensively consider the information of a plurality of leaf nodes. Phi is a_t(c ═ m | w) is defined as:

in the t-th intuitive fuzzy decision tree, B_tSet of all leaf nodes reached by the test sample w, b one leaf node reached by the test sample w, h^％(w) is the intuitive fuzzy membership degree of the test sample w to the current node with the leaf node b as the current node,the confidence that category m is predicted for leaf node b is defined as:

wherein x_jFor the training samples to reach leaf node b, there is n_bA 1, c_jFor training sample x_jIs a dirac function, h^％(x_j) Training sample x for leaf node b as current node_jAnd (4) intuitive fuzzy membership degree of the current node.

The training sample and the test sample are the same in calculation mode of the intuitive fuzzy membership degree of the current node, and the sample x is the intuitive fuzzy membership degree h of the current node^％(x) For all the branches it belongs to reach the current nodeThe product of the intuitive fuzzy membership of the output path of the point is specifically defined as:

where D is the set of all branch nodes that the sample passed before reaching the current node, D is one branch node in the set, l represents the output left branch of the branch node, r represents the output right branch of the branch node,the intuitive fuzzy membership degree of the output path which is subordinate to the branch node d and is passed by the sample to reach the current node.

The samples comprise test samples and training samples, the training samples comprise positive training samples and negative training samples, the positive training samples refer to the training samples in the target category, and the negative training samples refer to the training samples in the non-target category. If the current node is the root node, D is null, and h cannot be calculated by using the formula (4)^％(x) In that respect In this case, when the sample x is a test sample, h^％(x) When the sample x is a positive training sample and the total number of positive training samples is n ═ 1₁And n is₁When it is a positive integer, h^％(x)＝1/n₁When the sample x is a negative training sample and the total number of the negative training samples is n₀And n is₀When it is a positive integer, h^％(x)＝1/n₀。

Intuitive fuzzy membership of output paths subordinate to branch node d and through which sample reaches current nodeIs defined as:

whereinMembership of samples to tributary nodesThe output left branch of point d is the intuitive fuzzy membership,is the intuitive fuzzy membership of the output right branch whose sample is subordinate to the branch node d. When calculating according to equation (4), the output path of branch node d through which the sample should pass should be left branch or right branch according to equation (5)Andselects a corresponding one of the substitutes. h (x)_d) A decision function is output for the intuitive fuzzy of the branch node d.

The branch nodes of the traditional binary decision tree adopt hard decisions, and the definition of the output decision function of the branch nodes is as follows:

wherein x is_dIs the eigenvalue of the sample x of the branch node d, τ is the eigenvalue threshold. 0 corresponds to the branch node output left branch and 1 corresponds to the branch node output right branch. The conventional hard decision function represented by equation (19) is obfuscated using an S-type function (i.e., Sigmoid function).

Wherein x_dIs the characteristic value of the sample x of the branch node d, tau is the characteristic threshold value, theta is a constant parameter for controlling the degree of inclination of the Sigmoid function, and sigma is the standard deviation of the characteristic value.

By way of example, x_dHas a value range of [0,1 ]]Fig. 2 shows hard decision functions and fuzzy decision functions before and after blurring when τ is 0.4 and θ is 0.25. The dashed line in the figure represents the hard decision function defined by equation (19), the output of which jumps at the characteristic threshold; the solid line represents that defined by the formula (7)A fuzzy decision function whose output varies monotonically and continuously according to the eigenvalue of the sample, and is equal to 0.5 at the eigenvalue threshold.

And then, an intuitive fuzzy point operator is adopted, and a fuzzy decision function based on the Sigmoid function is further popularized to the intuitive fuzzy decision function.

Assuming that U is a non-empty set, the intuitive fuzzy set (IFS (U)) A of the set U is defined as:

A＝{<u,μ_A(u),ν_A(u)>|u∈U} (20)

wherein mu_A:U→[0,1]，μ_A(U) membership, ν, representing the membership of element U in set U to A_A:U→[0,1]，ν_A(U) represents the non-membership of the element U in the set U to A, and for any U:

the fuzzy intuitive index that the element U in the set U belongs to A is defined as:

π_A(u)＝1-μ_A(u)-ν_A(u) (22)

fuzzy intuition exponent pi_A(u) represents the uncertainty information of element u with respect to the set of intuitive ambiguities A. If pi_A(u) is small, indicating that the membership value of element u belonging to A is relatively accurate; if pi_AIf the value of (u) is large, it indicates that the membership value of element u belonging to A has a large uncertainty. Compared with a fuzzy set, the intuitive fuzzy set can embody the information of three aspects of membership, non-membership and fuzzy intuitive index, thereby being beneficial to better processing the uncertain information.

In order to better utilize the information in the fuzzy intuition index, an intuition fuzzy point operator is introduced. For any U ∈ U, let α_u,β_u∈[0,1]And satisfy α_u+β_uLess than or equal to 1, intuitive fuzzy point operatorIfs (u) → ifs (u) is defined as:

intuitive fuzzy point operatorConverting the intuitionistic fuzzy set A into an intuitionistic fuzzy set with fuzzy intuitionistic indexes as follows:

note the bookThen there are:

by analogy, it can be obtained for any positive integer n ifα_u+β_uNot equal to 0, then:

if for a certain U ∈ U, α_u+β_u0, i.e. alpha_u0 and beta_uWhen 0, then:

as can be seen from the expressions (27) and (28), the intuitive modeFuzzy point calculatorWill blur the intuitive index pi_A(u) is divided into: (1-. alpha.) with a high degree of polymerization_u-β_u)ⁿπ_A(u)，α_uπ_A(u)(1-(1-α_u-β_u)ⁿ)/(α_u+β_u) And beta_uπ_A(u)(1-(1-α_u-β_u)ⁿ)/(α_u+β_u) And the three parts respectively represent unknown, membership and non-membership parts in the original uncertain information.

For any U ∈ U, since α_u,β_u∈[0,1]And satisfy α_u+β_uLess than or equal to 1, comprising:

equation (30) shows that the intuitive fuzzy point operatorThe blur intuitive index of the intuitive blur set a can be reduced. This illustrates the operator of the fuzzy point through intuitionNew information can be extracted from the uncertain information of the element u relative to the intuitive fuzzy set A, and the utilization degree of the uncertain information is improved.

Intuitive fuzzy output decision function h (x) of branch node d obtained by intuitive fuzzy popularization_d) Is defined as:

and k is the operator times and is a positive integer, and the larger the value of k is, the larger the consumed calculated amount is. For computational convenience, k ∈ {1,2,3} can be taken.

Pi (z) in equation (6) is a fuzzy intuitive index, defined according to Sugeno fuzzy complement as:

where λ is a constant parameter, 0<λ<1 is, for example, 0.8. When x is_dWhen t is greater than or equal to z, g (x)_d) When x is_d<τ, z is 1-g (x)_d)。

In the formula (6), α is a scale factor for extracting membership information from the fuzzy intuitive index, and β is a scale factor for extracting non-membership information from the fuzzy intuitive index, and is defined as:

β＝1-α-π(z)

the output value of equation (6) represents the intuitive fuzzy membership of the sample to the right branch of the branch node output. Intuitive fuzzy point operatorCan extract new useful information from the uncertain information, and the formula (6) is the information g (x) of the original fuzzy membership degree_d) Membership information extracted from the fuzzy intuitive index pi (z) is added, so that the uncertainty of the original fuzzy membership information is reduced. Since when k is 0, h (x) can be obtained from formula (6)_d)＝g(x_d) At this time, the intuitive fuzzy decision is degraded to a fuzzy decision, and therefore, it can be considered that expression (6) is an intuitive fuzzy generalization to expression (7).

For example, when k is 1, the characteristic threshold τ is 0.4, and λ is 0.8, the graph of the branch node intuitively blurs the output decision function is shown in fig. 3. The dotted line in the figure represents the fuzzy decision function defined by equation (7); the solid line represents the intuitive fuzzy decision function defined by equation (6).

The values of the operator times k and the characteristic threshold value tau in the formula (6) can be determined by training the characteristic selection criterion in the process of updating the intuitive fuzzy decision tree, and can also be determined by other modes such as empirical values.

The branch nodes of the traditional binary decision tree adopt hard decision, a test sample can only select one from the left branch and the right branch to reach the next layer of nodes according to characteristic attributes, and finally reaches one leaf node, and the category of the test sample is determined by the category of the reached leaf node. The robustness of the hard decision tree to the noise of the sample is not strong, when the sample is interfered by strong noise, the characteristic value of the hard decision tree is greatly changed, so that the branch through which the sample passes is possibly changed, and the accuracy of the decision tree is reduced.

In the prior art, a fuzzy decision tree is provided, a fuzzy set theory is applied to the training and reasoning process of the decision tree, and the representation capability of the fuzzy set theory is utilized to improve the processing capability of the traditional decision tree on noisy data and incomplete data. Although the fuzzy decision tree can process the characteristic value with uncertainty, the fuzzy semantic processing needs to be carried out on the sample characteristic, and the sample characteristic adopted in the target tracking is mostly numerical type characteristic and the characteristic dimension is high, so that the fuzzy semantic processing on the sample characteristic becomes difficult.

In the embodiment, the classification result of the test sample is obtained by comprehensively considering the intuitionistic fuzzy membership degree calculated by each intuitionistic fuzzy decision tree, and the classification performance of the intuitionistic fuzzy random forest is better than that of a single intuitionistic fuzzy decision tree. The intuitive fuzzy random tree fuzzifies hard decision of a traditional decision tree by adopting a Sigmoid function, omits a complex fuzzy semantic process, and adopts an intuitive fuzzy point operator to popularize the fuzzy membership degree to the intuitive fuzzy membership degree, so that useful information is extracted, and the robustness is improved.

As shown in fig. 4, a third embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention is based on the first embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention, wherein updating the intuitive fuzzy random forest for the reliable target successfully associated or successfully matched includes:

s41: and updating the training sample set.

And taking a target image block corresponding to the successfully associated or successfully matched reliable target in the current video frame as a new positive training sample to be added into the positive training sample set, selecting a plurality of image blocks in a specified range around the positive training sample as negative training samples, and forming a training sample set by the positive training sample set and the negative training samples. The positive training sample set in this embodiment may include all corresponding image blocks of the reliable target in the current and previous video frames, and may also limit the number of positive training samples in the positive training sample set to be less than or equal to a specified threshold value to save storage resources.

S42: a number of samples are recursively randomly sampled from a set of training samples to obtain a subset of training samples.

This means that there may be repeated occurrences of the same training sample in the subset of training samples, and overfitting may be avoided.

S43: and generating an intuitive fuzzy decision tree by using the training sample subset.

S44: and judging whether the number of the generated intuitive fuzzy decision trees reaches a preset number T.

If yes, ending the process, and forming a new intuitive fuzzy random forest by the generated T intuitive fuzzy decision trees; if not, the step S42 is returned to continue the circulation.

This embodiment may be combined with the second embodiment of the target tracking method based on the intuitive fuzzy random forest of the present invention.

As shown in fig. 5, a fourth embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention is based on the third embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention, and step S43 specifically includes:

s431: and initializing intuitive fuzzy membership degrees of the training samples in the training sample subset to the root node.

Total number n of positive training samples in the subset of training samples₁When the training sample is a positive integer, the intuitive fuzzy membership degree of the positive training sample to the root node is 1/n₁Total number n of negative training samples in the subset of training samples₀When the training sample is a positive integer, the intuitive fuzzy membership degree of the negative training sample to the root node is 1/n₀。

S432: and training the feature selection criterion of the training samples reaching the current node.

The initial current node is the root node.

And confirming the optimal one-dimensional characteristic of the current node, the operator times of the optimal one-dimensional characteristic and the value of a characteristic threshold value according to the intuitionistic fuzzy information gain maximum principle, wherein the optimal one-dimensional characteristic belongs to a high-dimensional characteristic vector of a training sample.

S433: and judging whether the current node meets the stop condition.

The deeper the depth of the intuitive fuzzy decision tree, the more memory resources it consumes, and the more computation is required, and therefore, it is necessary to design a stop condition for its generation process. The stop condition may include:

1) the proportion of the sum of the intuitive fuzzy membership degrees of the training samples belonging to a certain class and reaching the current node to the sum of the intuitive fuzzy membership degrees of all the training samples reaching the current node is greater than a third threshold value theta_r；

2) The sum of the intuitive fuzzy membership degrees of the training samples reaching the current node and belonging to the current node is less than a fourth threshold value theta_l；

3) The depth of the current node in the intuitive fuzzy decision tree reaches a fifth threshold value theta_d。

If any of the above three conditions is satisfied, the process proceeds to step S434, and if none of the conditions is satisfied, the process proceeds to step S435.

S434: and converting the current node into a leaf node.

S435: and splitting the current node by using the optimal one-dimensional characteristics to generate two branch nodes of the next layer.

And then, taking the branch node as the current node, returning to the step S432, and continuing to execute until all current nodes become leaf nodes and no branch node is generated. The generation process of the intuitive fuzzy decision tree is a process of recursively constructing a binary tree by starting from a root node and taking the maximization of the intuitive fuzzy information as a characteristic selection criterion.

As shown in fig. 6, step S432 specifically includes:

s410: a one-dimensional feature is randomly selected from the high-dimensional feature vector of the training sample.

S420: selecting one of the candidate feature threshold values, calculating the intuitive fuzzy information gain when the operator times take different values under the condition of the selected one-dimensional feature and the feature threshold value, and recording the selected one-dimensional feature, the value of the feature threshold value, the maximum intuitive fuzzy information gain and the value of the corresponding operator times.

In an embodiment of the present invention, the candidate feature threshold may include a median of two adjacent values obtained by sorting the values of the selected one-dimensional features of the training samples, and n training samples may obtain n-1 medians. The candidate feature threshold may also include an average of values of selected one-dimensional features of all training samples. Of course, a combination of the two may be used.

The intuitive blur information gain Δ H is defined as:

wherein X ═ { X ═ X₁,x₂,...,x_nAnd n is the number of training samples in the set X. In general, the training samples can reach each node in the intuitive fuzzy decision tree through each output path of each branch node, so X is the initialized training sample subset.

H (X) is the intuitive fuzzy entropy of set X, defined as:

where δ (·) is a dirac function, c_jTo train class labels of samples, m_iFor training the class of the sample, since only the target and the non-target need to be distinguished, there are two classes, i is 1 and 2. Intuitive fuzzy membership h of training sample to current node^％(x_j) The definitions and calculation manners of (a) can refer to expressions (4) - (9), it should be noted that the sample X in expressions (4) - (9) is the training sample in the set X,the sample features, feature threshold values and operator times used are of the branch node before the current node is reached.

H^l(X) the intuitive fuzzy entropy of the set of training samples contained in the left branch for the current node output is defined as:

H^r(X) is the intuitive fuzzy entropy of the set of training samples contained in the right branch output by the current node, defined as:

whereinIs the intuitive fuzzy membership of the output left branch whose sample is subordinate to the current node,the intuitive fuzzy membership degree of the output right branch of the current node to which the sample belongs is calculated in the same manner as that of the branch node d, and equations (5) - (9) can be referred to, and it should be noted that the sample X in equations (5) - (9) is still the training sample in the set X at this time, but the selected one-dimensional feature and the feature threshold value of the current node and the operator number specified this time are used.

And respectively calculating the gain delta H of the intuitive fuzzy information when the operator times take different values, and finding out the maximum gain delta H of the intuitive fuzzy information from the gain delta H for recording.

S430: the previous step (i.e., step S420) is performed for each of the candidate feature threshold values to find and store the one of all records with the largest gain of the intuitive fuzzy information.

The one-dimensional feature included in the record is the optimal one-dimensional feature, and the value of the feature threshold value and the value of the operator times are the operator times and the value of the feature threshold value of the optimal one-dimensional feature.

As shown in fig. 7, a fifth embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention is based on the third embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention, and step S43 specifically includes:

s436: and initializing intuitive fuzzy membership degrees of the training samples in the training sample subset to the root node.

S437: and judging whether the current node meets the stop condition.

The initial current node is the root node.

If any of the above three conditions is satisfied, the process proceeds to step S438, and if none of the conditions is satisfied, the process proceeds to step S439.

S438: and converting the current node into a leaf node.

S439: and training a feature selection criterion on a training sample reaching the current node, and splitting the current node by using the optimal one-dimensional features to generate two branch nodes of the next layer.

Then, the branch nodes are used as current nodes and returned to step S437 to continue execution until all current nodes become leaf nodes and no branch nodes are generated.

The difference between this embodiment and the fourth embodiment of the target tracking method based on the intuitive fuzzy random forest of the present invention is that the execution sequence of the step of determining whether the current node meets the stopping condition is different from the execution sequence of the step of training the feature selection criterion on the training sample reaching the current node, and specific contents may refer to the fourth embodiment of the target tracking method based on the intuitive fuzzy random forest of the present invention, and are not described herein again.

As shown in fig. 8, a sixth embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention is based on the first embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention, and step S2 includes:

s21: a similarity measure between the observed and predicted outcomes is calculated.

The similarity measures include a spatial distance feature similarity measure and an appearance feature similarity measure.

In general, the position of the target between adjacent frame images does not change greatly, and therefore, the spatial distance feature is one of features that can more effectively match the observed result and predicted result of the target. Spatial distance feature similarity metric ψ between observation d and prediction o₁Is defined as:

wherein | · | purple₂Is a two-norm, (xo, yo) is the center coordinate of the predicted result o, (x)_d,y_d) As the center coordinate of observation d, h_oIn order to predict the height of the result o,as a constant of variance, can be taken

Since the appearance of an object may change over time, a single fixed object template may not form an accurate description of the appearance of the object, and therefore, a set of object templates is used to represent the appearance of the object. The target template set corresponding to the prediction result o isWherein the target template e_i,i＝1,...,n₂Is the first n subjected to whitening processing and scaled to h × w₂Associated/matching object image blocks in individual video frames, n₂Is the total number of target templates included in the set of target templates. For storage and computational convenience, the number of target templates contained in the set of target templates is limited, n₂Less than or equal to the seventh threshold γ, γ may be equal to 5.

Appearance feature similarity measure ψ between observed result d and predicted result o₂Is defined as:

wherein s (-) is the observation d and the target template e_iA normalized correlation metric between, defined as:

where d (x, y) is the gray scale value of the observation d at the coordinates (x, y), e_i(x, y) as a target template e_iThe gray value at coordinate (x, y) and observation d is also whitened and scaled in size to h x w. s has a value range of [0,1 ]]。

S22: and calculating the correlation cost between the observed result and the predicted result by using the similarity measurement.

The cost of association between observed outcome d and predicted outcome o is defined as:

ρ_o,d＝1-ψ₁×ψ₂ (17)

s23: and calculating an optimal incidence matrix between the observed result and the predicted result by utilizing the incidence cost as an incidence result.

The set of all observations is D ═ D₁,...,d_pAll the prediction results are combined into a set of O ═ O }₁,...,o_qAnd the total associated cost of the observed result and the predicted result is defined as:

where ρ is_ijObservation d defined for formula (17)_iAnd the predicted result o_jAssociated cost between, a ═ a_ij]_p×qIs a correlation matrix between observed results and predicted results, any element a in the correlation matrix_ijE {0,1}, when a_ijWhen 1, the observation result d is shown_iAnd the predicted result o_jThe association is successful.

Solving since one observation can only be associated with one object and one object can also be associated with only one observationObtaining a correlation matrix A which minimizes the total correlation cost of the observed result and the predicted result₀I.e. the optimal correlation matrix. The correlation result can be obtained by solving using the Hungarian algorithm.

As shown in fig. 9, a seventh embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention is based on the sixth embodiment of the target tracking method based on an intuitive fuzzy random forest according to the present invention, and after step S3, the method further includes:

s5: and updating the target template set for the reliable targets with successful association or successful matching.

And (3) carrying out whitening processing on the image block of the association/matching object successfully associated or matched with the current frame of the reliable target, and adding the image block of the association/matching object into the target template set of the reliable target after the size of the image block is scaled to h multiplied by w. And if the number of the target templates in the target template set before the target templates are added is equal to the seventh threshold value, deleting the target template which is added in the target template set earliest.

This step and step S4 may be executed independently of each other or simultaneously.

The following is a result of experimental verification and comparison performed by using an embodiment of the target tracking method based on the intuitive fuzzy random forest, which is a combination of the first to fourth, sixth and seventh embodiments of the present invention, and is a video multi-target tracking (IFRFMOT) algorithm based on the intuitive fuzzy random forest, in which a kalman filter is used to filter and predict an effective target track and a temporary target track. The model parameters selected by the IFRFMOT algorithm of this embodiment include: the size of the target template is 64 × 32. The intuitive fuzzy random forest consists of 10 intuitive fuzzy decision trees, and the maximum depth of each intuitive fuzzy decision tree is 7. The intuitive fuzzy random forest uses RGB color channel images as sample features.

The experimental subject adopts 2 representative public test videos TownCentre and PETS.S2L2, and the 2 test videos are monitoring videos in a public scene. In order to comprehensively and accurately evaluate the Tracking performance of the IFRFMOT algorithm of this embodiment, 6 commonly used Tracking performance evaluation indexes are adopted, namely, target tag change times (IDS), multi-target Tracking Accuracy (MOTA), multi-target Tracking Precision (MOTP), target proportion for long-time Tracking (MT), target proportion for short-time Tracking (ML), and target track disconnection times (Fragmentation), where the larger the values of MOTA, MOTP, and MT are, the better the Tracking performance is, and the smaller the values of IDS, ML, and FG are, the better the Tracking performance is. This embodiment implements the same decision-making of the video-based random decision-making algorithm (more than one) as the TC _ ODAL algorithm (more than one S H, bamboo KJ. routing on tracking less control and linking estimation left [ C ]. in IEEE Conference on Computer Vision and Pattern registration, Columbus, OH,2014: 1218-.

The experimental results of the IFRFMOT algorithm of this example and the comparison algorithm for the test video TownCentre are shown in table 1.

TABLE 1

As can be seen from table 1, the IFRFMOT algorithm is significantly better than the comparison algorithm in the MOTA index, which indicates that the IFRFMOT algorithm has fewer tracking errors than the comparison algorithm. Compared with the TC _ ODAL algorithm, the IFRFMOT algorithm is improved by 39.1% on the ML index and 16.7% on the MT index; compared with the MDP algorithm, the IFRFMOT algorithm is improved by 6.2% on the ML index and 1.3% on the MT index; compared with the MHT algorithm, the IFRFMOT algorithm improves the ML index by 14.1 percent and achieves the same accuracy rate on the MT index. The IFRFMOT algorithm is proved to have more accurately tracked targets and complete output target tracks. Compared with the RFMOT algorithm, the IFRFMOT algorithm is improved by 0.9 percent on the ML index, is improved by 0.4 percent on the MT index and is reduced by 9 times of target label distribution errors on the IDS index. This result further verifies the effectiveness of the intuitive fuzzy random forest. Under the influence of high-frequency occlusion and target track crossing, the IFRFMOT algorithm is inferior to the TC _ ODAL algorithm, the MDP algorithm and the MHT algorithm in IDS indexes, but the IFRFMOT algorithm estimates 60% of target tracks, and the number of targets accurately tracked by the IFRFMOT algorithm at the same time is more than that of the 3 comparison algorithms. Due to the fact that the number of training samples is small, although the IFRFMOT algorithm can track the undetected target on line, the candidate result is not accurately positioned, the filtering effect of an adopted Kalman filter is reduced due to the fact that the target moves nonlinearly, and the IFRFMOT algorithm is inferior to a comparison algorithm in FG indexes.

S2l2, the IFRFMOT algorithm and the comparison algorithm of this embodiment are shown in table 2.

TABLE 2

As can be seen from table 2, the IFRFMOT algorithm is significantly better than the comparison algorithm in the MOTA index, which indicates that the IFRFMOT algorithm has fewer tracking errors than the comparison algorithm. Also, the IFRFMOT algorithm is significantly better than the comparison algorithm in ML metric. Compared with the TC _ ODAL algorithm, the IFRFMOT algorithm achieves 20.9% improvement on the ML index, and the target track accurately tracked by the IFRFMOT algorithm is more than that tracked by the comparison algorithm. Meanwhile, compared with the TC _ ODAL algorithm, the IFRFMOT algorithm is improved by 16.3% on the MT index; compared with the MDP algorithm, the IFRFMOT algorithm is improved by 9.3% on the MT index; compared with the MHT algorithm, the IFRFMOT algorithm is slightly lower than the MT index by 2.3 percent, which shows that the IFRFMOT algorithm not only has more accurately tracked targets, but also has more complete output target tracks. Compared with the RFMOT algorithm, the IFRFMOT algorithm is improved by 2.3 percent on the ML index, is improved by 2.3 percent on the MT index and is reduced by 15 times of target label distribution errors on the IDS index. This result further verifies the validity of the intuitive fuzzy random forest model. S2l2, is inferior to MHT, but is superior to other contrast algorithms in IDS and FG indicators.

An IFRFMOT algorithm is realized by adopting an MATLAB programming language, and an experimental platform is a desktop computer with an Intel dual-core 3.6GHz processor and an 8GB memory. The average processing speed of the IFRFMOT algorithm and the comparison algorithm on the test video is shown in table 3. The processing time consumed for motion detection has been excluded from the statistics of table 3.

TABLE 3

As can be seen from table 3, the IFRFMOT algorithm is slightly inferior to the comparison algorithm in terms of operation efficiency, but since the training process of each intuitive fuzzy decision tree in the intuitive fuzzy random forest is independent from each other, the operation process can be accelerated by parallelization, and therefore, if the IFRFMOT algorithm is implemented by using more efficient C language programming and parallel computation is performed by using multiple CPUs or GPUs, the operation efficiency of the IFRFMOT algorithm can be further improved.

Aiming at a test video TownCentre with numerous target tracks and strong background interference and a test video PETS.S2L2 with large target density and obvious illumination change, the IFRFMOT algorithm has better or close tracking performance to a contrast algorithm on the aspects of multi-target tracking accuracy, target proportion tracked for a long time and target proportion tracked for a short time, mainly because the IFRFMOT algorithm populates a random forest based on hard decision to be intuitive fuzzy random due to background interference and illumination change and a large amount of target missing detection caused by high frequency shielding among targets, and can better process the uncertainty of feature description and effectively distinguish different targets and targets from the background. Therefore, even if the originally stably tracked target is missed, the IFRFMOT algorithm of the embodiment can still track the target online by using the intuitive fuzzy random forest.

As shown in fig. 10, the first embodiment of the target tracking apparatus based on the intuitive fuzzy random forest according to the present invention includes:

and the detection module 11 is configured to perform motion detection on the current video frame, and obtain a possible motion object as an observation result.

And the association module 12 is configured to associate the observation result with a prediction result of the target, where the prediction result is obtained by predicting at least a trajectory of the target in a previous video frame, and the target includes a reliable target and a temporary target.

And the management module 13 is configured to perform trajectory management on the observation result and the prediction result that are not associated, where the trajectory management includes performing online tracking on the prediction result of the reliable target that is not associated to obtain a candidate result, and matching the candidate result by using an intuitive fuzzy random forest of the reliable target that is not associated.

An updating module 14, configured to obtain a trajectory of the target of the current frame by using the association result and the matching result, where the trajectory is obtained by filtering and updating a prediction result of a successfully matched reliable target by using a successfully matched candidate result; and predicting by utilizing the track of the target of the current frame, and updating the intuitive fuzzy random forest for the successfully associated or matched reliable target.

As shown in fig. 11, the second embodiment of the target tracking device based on the intuitive fuzzy random forest according to the present invention comprises: a processor 110 and a camera 120. The camera 120 may be a local camera, and the processor 110 is connected to the camera 120 through a bus; the camera 120 may also be a remote camera and the processor 110 may be connected to the camera 120 via a local area network or the internet.

The processor 110 controls the operation of the target tracking device based on an intuitive fuzzy random forest, and the processor 110 may also be referred to as a Central Processing Unit (CPU). The processor 110 may be an integrated circuit chip having signal processing capabilities. The processor 110 may also be a general purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The target tracking device based on the intuitive fuzzy random forest may further include a memory (not shown) for storing instructions and data necessary for the processor 110 to operate and also storing video data photographed by the transmitter 120.

The processor 110 is configured to perform motion detection on a current video frame acquired from the camera 120, and detect a possible motion object as an observation result; correlating the observation result with a prediction result of the target, wherein the prediction result is obtained by predicting at least the track of the target of the previous video frame, and the target comprises a reliable target and a temporary target; performing track management on the observation result and the prediction result which are not associated, wherein the track management comprises the steps of performing online tracking on the prediction result of the reliable target which is not associated to obtain a candidate result, and matching the candidate result by utilizing an intuitionistic fuzzy random forest of the reliable target which is not associated; obtaining the track of the target of the current frame by using the correlation result and the matching result, wherein the track is obtained by filtering and updating the prediction result of the successfully matched reliable target by using the successfully matched candidate result; and predicting by utilizing the track of the target of the current frame, and updating the intuitive fuzzy random forest for the successfully associated or matched reliable target.

The functions of each part included in the target tracking device based on the intuitive fuzzy random forest can refer to the description in each corresponding embodiment of the online target tracking method, and are not described in detail herein.

The above description is only an embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes performed by the present specification and drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims

1. A target tracking method based on an intuitionistic fuzzy random forest is characterized by comprising the following steps:

performing motion detection on the current video frame, wherein a possible motion object obtained by detection is used as an observation result;

correlating the observation result with a prediction result of a target, wherein the prediction result is obtained by predicting at least the track of the target of a previous video frame, and the target comprises a reliable target and a temporary target;

performing trajectory management on the observation result and the prediction result which are not associated, wherein the trajectory management comprises performing online tracking on the prediction result of the reliable target which is not associated to obtain a candidate result, and matching the candidate result by using an intuitive fuzzy random forest of the reliable target which is not associated to;

obtaining the track of the target of the current frame by using the correlation result and the matching result, wherein the track is obtained by filtering and updating the prediction result of the reliable target which is successfully matched by using the candidate result which is successfully matched by the reliable target; predicting by utilizing the track of the target of the current frame, and updating the intuitive fuzzy random forest for the reliable target which is successfully associated or matched;

wherein the online tracking the prediction result of the reliable target which is not associated to obtain the candidate result comprises:

selecting a number of image blocks within a specified range around and at the predicted outcome location of the reliable target as the candidates, which may include the observations not associated;

said matching said candidate results with an intuitive fuzzy random forest of said reliable targets not associated comprising:

calculating the candidate result by using the intuitive fuzzy random forest as the intuitive fuzzy membership degree of the test sample to the reliable target class;

if the intuitionistic fuzzy membership is greater than a first threshold value and the similarity measure of the appearance characteristics of the candidate result and the prediction result of the reliable target is greater than a second threshold value, the matching is successful;

wherein the intuitive fuzzy degree of membership P (c m | w) of the candidate result as a test sample to the reliable target class is:

wherein c is a class label of the test sample, M is the reliable target class, w is the test sample, T is the number of intuitive fuzzy decision trees in the intuitive fuzzy random forest, M is a class set of training samples for generating the intuitive fuzzy random forest, phi_t(c ═ m | w) is the intuitive fuzzy membership degree of the test sample w to the reliable target class m calculated by using the t-th intuitive fuzzy decision tree, and is defined as:

in the t-th of the intuitive fuzzy decision tree, B_tA set of all leaf nodes reached by the test sample w, b a leaf node reached by the test sample w, h^％(w) is the intuitive fuzzy degree of membership of the test sample w to the current node with the leaf node b as the current node,predicting the confidence coefficient of the category m for the leaf node b, and defining as follows:

wherein x_jFor the training samples to reach the leaf node b, there is n_bA 1, c_jIs the training sampleThis x_jIs a dirac function, h^％(x_j) The training sample x is the current node of the leaf node b_j(ii) an intuitive fuzzy degree of membership to the current node;

intuitive fuzzy membership definition h where sample x is attached to the current node^％(x) Is defined as:

the samples comprise the test samples and the training samples, the training samples comprise positive training samples and negative training samples, if the current node is a root node, when the sample x is the test sample, h^％(x) When the sample x is the positive training sample and the total number of the positive training samples is n ═ 1₁And n is₁When it is a positive integer, h^％(x)＝1/n₁When the sample x is the negative training sample and the total number of the negative training samples is n₀And n is₀When it is a positive integer, h^％(x)＝1/n₀；

Where D is the set of all branch nodes that the sample passed before reaching the current node, D is one branch node in the set, l represents the output left branch of the branch node, r represents the output right branch of the branch node,the intuitive fuzzy membership degree of the output path which is subordinate to the branch node d and is passed by the sample to reach the current node is defined as:

whereinMembership of the sample to the branch nodeThe output left branch of point d is the intuitive fuzzy membership,an intuitive fuzzy degree of membership, h (x), for the output right branch of the sample to the branch node d_d) Outputting a decision function for the intuitive fuzzy of the branch node d, which is defined as:

wherein x_dK is the operator number and is a positive integer, and is the characteristic value of the sample of the branch node;

g(x_d) Is an S-type function and is defined as:

wherein τ is a characteristic threshold value, θ is a constant parameter for controlling the degree of tilt of the sigmoid function, and σ is a standard deviation of the characteristic value;

π (. cndot.) is a fuzzy intuitive index defined as:

where λ is a constant parameter, 0<λ<1, when x_dWhen t is greater than or equal to z, g (x)_d) When x is_d<τ, z is 1-g (x)_d)；

Alpha is a scale factor for extracting membership information from the fuzzy intuitive index, beta is a scale factor for extracting non-membership information from the fuzzy intuitive index, and is defined as follows:

and the values of the operator times k and the characteristic threshold value tau are determined by training the characteristic selection criterion in the process of updating the intuitive fuzzy decision tree.

2. The method of claim 1,

the updating the intuitive fuzzy random forest for the reliable target with successful association or successful matching comprises:

taking a target image block corresponding to the reliable target successfully associated or matched in the current video frame as a new positive training sample to be added into a positive training sample set, and selecting a plurality of image blocks as negative training samples in a specified range around the positive training sample;

randomly sampling a plurality of samples in a replacing way from the training sample set consisting of the positive training sample set and the negative training sample set to obtain a training sample subset;

generating an intuitive fuzzy decision tree using the subset of training samples;

and circularly executing the steps to generate a preset number of the intuitive fuzzy decision trees to form the new intuitive fuzzy random forest.

3. The method of claim 2,

the generating an intuitive fuzzy decision tree using the subset of training samples comprises:

initializing an intuitive fuzzy membership degree of training samples in the training sample subset to a root node, wherein the total number n of positive training samples in the training sample subset₁When the training sample is a positive integer, the intuitive fuzzy membership degree of the positive training sample to the root node is 1/n₁Total number n of negative training samples in the subset of training samples₀When the training sample is a positive integer, the intuitive fuzzy membership degree of the positive training sample to the root node is 1/n₀；

Performing feature selection criterion training on the training sample reaching the current node, confirming the optimal one-dimensional feature of the current node, the operator times of the optimal one-dimensional feature and the value of a feature threshold according to an intuitionistic fuzzy information gain maximum principle, judging whether the current node meets a stop condition or not, if so, converting the current node into a leaf node, and if not, splitting the current node by using the optimal one-dimensional feature to generate two branch nodes of a next layer; or judging whether the current node meets a stopping condition, if so, converting the current node into a leaf node, if not, performing the feature selection criterion training on the training sample reaching the current node, and then splitting the current node by using the optimal one-dimensional feature to generate two branch nodes of a next layer;

and taking the branch node as the current node, returning to the previous step and continuing to execute.

4. The method of claim 3,

the training the feature selection criterion on the training samples reaching the current node comprises:

randomly selecting one-dimensional feature from the high-dimensional feature vector of the training sample;

selecting one of candidate feature threshold values, calculating the intuitive fuzzy information gain when the operator times take different numerical values under the condition of the selected one-dimensional feature and the feature threshold value, and recording the selected one-dimensional feature, the value of the feature threshold value, the maximum intuitive fuzzy information gain and the corresponding value of the operator times;

executing the previous step for each candidate characteristic threshold value, and finding and storing the piece with the largest gain of the intuitive fuzzy information in all records;

and repeatedly executing the steps for specified times, and finding out the piece with the largest gain of the intuitive fuzzy information in all the obtained stored records, wherein the included one-dimensional characteristic is the optimal one-dimensional characteristic, and the value of the characteristic threshold value and the value of the operator times are the operator times and the value of the characteristic threshold value of the optimal one-dimensional characteristic.

5. The method of claim 4,

the intuitive blur information gain Δ H is defined as:

wherein X ═ { X ═ X₁,x₂,...,x_nThe n is the number of training samples in the set;

h (X) is the intuitive fuzzy entropy of the set X, defined as:

where δ (·) is a dirac function, c_jIs a class label of the training sample, m_iFor the category of the training sample, since only the classification of belonging to the target and not belonging to the target is required, there are two categories, i is 1, 2;

training sample membership intuitive fuzzy membership h to the current node^％(. cndot.) is defined as:

where D is the set of all branch nodes that the sample passes through before reaching the current node, D is one branch node in the set,the intuitive fuzzy membership of the output left branch for which the sample is subordinate to the branch node d,the intuitive fuzzy membership degree of the output right branch of the sample belonging to the branch node d is defined as:

wherein h (x)_d) Outputting a decision function for the intuitive fuzzy of the branch node d, which is defined as:

wherein x_dThe characteristic value of the training sample of the branch node is obtained, k is the operator frequency and is a positive integer;

g(x_d) Is an S-type function and is defined as:

wherein τ is the characteristic threshold value, θ is a constant parameter for controlling the degree of tilt of the sigmoid function, and σ is a standard deviation of the characteristic value;

π (. cndot.) is a fuzzy intuitive index defined as:

H^l(X) the intuitive fuzzy entropy of the set of training samples contained in the output left branch for the current node is defined as:

H^r(X) is the intuitive fuzzy entropy of the set of training samples contained by the current node output right branch, defined as:

whereinAn intuitive fuzzy degree of membership of the sample to the output left branch of the current node,and calculating the intuitive fuzzy membership degree of the output right branch of the current node to which the sample belongs in the same way as the branch node d, wherein the used parameters are the selected one-dimensional characteristic, the characteristic threshold value and the operator times.

6. The method according to claim 4, wherein the candidate feature threshold comprises a median of two adjacent values obtained by sorting the values of the selected one-dimensional features of the training samples, and/or an average of the values of the selected one-dimensional features of all the training samples.

7. The method of claim 3,

the stop condition includes:

the proportion of the sum of the intuitive fuzzy membership degrees of the training samples belonging to a certain class reaching the current node and the intuitive fuzzy membership degree of all the training samples reaching the current node is greater than a third threshold value;

or the sum of the intuitive fuzzy membership degrees of the training samples arriving at the current node and belonging to the current node is less than a fourth threshold value;

or the depth of the current node in the intuitive fuzzy decision tree reaches a fifth threshold.

8. The method according to any one of claims 1 to 7,

the trajectory managing the observations and the predictions that are not associated further comprises:

establishing a new temporary target for the observation result which is not associated and is not a candidate result of successful matching, changing the temporary target with the number of frames successfully associated larger than a first frame number threshold into a reliable target, deleting the temporary target with the number of frames unsuccessfully associated larger than a second frame number threshold, deleting the temporary target with the number of frames unsuccessfully associated larger than a third frame number threshold, and obtaining the matching result which is the reliable target with failed matching, wherein the step of obtaining the matching result which is the matching failure means that the intuition fuzzy membership degree of the candidate result as a test sample belonging to the reliable target category is calculated by using the intuition fuzzy random forest, and is smaller than or equal to a sixth threshold;

the obtaining of the track of the target of the current frame by using the correlation result and the matching result further includes:

and filtering and updating the prediction result of the target which is successfully associated by utilizing the associated observation result to obtain the track, taking the corresponding observation result as the track for the new temporary target, and taking the prediction result of the temporary target which is not successfully associated and not deleted and the reliable target which is not successfully associated and not matched as the track.

9. The method according to any one of claims 1 to 7,

the correlating the observation with the prediction of the target comprises:

calculating a similarity measure between the observation and the prediction, the similarity measure comprising a spatial distance feature similarity measure and an appearance feature similarity measure;

calculating an association cost between the observed outcome and the predicted outcome using the similarity measure;

and calculating an optimal incidence matrix between the observed result and the predicted result by using the incidence cost as an incidence result, so that the total incidence cost of the observed result and the predicted result is minimum.

10. The method of claim 9, comprising:

the spatial distance feature similarity measure ψ between observation d and prediction o₁Is defined as:

wherein | · | purple₂Is a two-norm, (x)_o,y_o) (x) is the center coordinate of the prediction result o_d,y_d) Is the center coordinate of the observation d, h_oIs the height of the prediction result o,is a constant of variance;

the target template set corresponding to the prediction result o isWherein the target template e_i,i＝1,...,n₂Is the first n subjected to whitening processing and scaled to h × w₂Associated/matching object image blocks in individual video frames, n₂The appearance feature similarity measure ψ between the observed result d and the predicted result o for the total number of the target templates included in the target template set and less than or equal to a seventh threshold value₂Is defined as:

where d (x, y) is the gray value of the observation d at the coordinates (x, y), e_i(x, y) is the target template e_iA gray value at coordinate (x, y);

the associated cost between the observed outcome d and the predicted outcome o is defined as:

ρ_o,d＝1-ψ₁×ψ₂ (17)

the set of all the observations is D ═ D₁,...,d_pAll the prediction results form a set of O ═ O }₁,...,o_q-a total associated cost of said observed and predicted outcomes is defined as:

where ρ is_ijAs an observation d_iAnd the predicted result o_jAssociated cost between, a ═ a_ij]_p×qIs a correlation matrix between the observed result and the predicted result, and any element a in the correlation matrix_ijE {0,1}, when a_ijWhen 1, the observation result d is expressed_iAnd the predicted result o_jThe association is successful; the optimal incidence matrix is solvedThe obtained incidence matrix A₀。

11. The method of claim 10, further comprising:

and for the reliable target which is successfully associated or successfully matched, adding the associated/matched object image block into the target template set of the reliable target after whitening and scaling the size to h multiplied by w, and deleting the target template which is added earliest in the target template set if the number of the target templates in the target template set before the reliable target is added is equal to the seventh threshold.

12. A target tracking device based on an intuitive fuzzy random forest is characterized by comprising: the processor is connected with the camera;

the processor is used for carrying out motion detection on the current video frame acquired from the camera, and a possible motion object obtained by detection is used as an observation result; correlating the observation result with a prediction result of a target, wherein the prediction result is obtained by predicting at least the track of the target of a previous video frame, and the target comprises a reliable target and a temporary target; trajectory management of the observations and the predictors not associated, comprising selecting a number of image blocks as candidates within a specified range around and at the position of the predictor of the reliable target not associated, which candidates may comprise the observations not associated, matching the candidates using an intuitive fuzzy random forest of the reliable target not associated, the matching of the candidates using an intuitive fuzzy random forest of the reliable target not associated comprising calculating the intuitive fuzzy membership of the candidates as test samples belonging to the class of reliable targets using the intuitive fuzzy random forest if the intuitive fuzzy membership is larger than a first threshold and the measure of similarity of appearance features of the candidates to the predictor of the reliable target is larger than a second threshold, matching is successful; obtaining the track of the target of the current frame by using the correlation result and the matching result, wherein the track is obtained by filtering and updating the prediction result of the reliable target which is successfully matched by using the candidate result which is successfully matched by the reliable target; predicting by utilizing the track of the target of the current frame, and updating the intuitive fuzzy random forest for the reliable target which is successfully associated or matched;

wherein x_jFor the training samples to reach the leaf node b, there is n_bA 1, c_jFor the training sample x_jIs a dirac function, h^％(x_j) The training sample x is the current node of the leaf node b_jSubject to the instituteIntuitive fuzzy membership of the current node;

whereinThe intuitive fuzzy membership of the output left branch for which the sample is subordinate to the branch node d,an intuitive fuzzy degree of membership, h (x), for the output right branch of the sample to the branch node d_d) Outputting a decision function for the intuitive fuzzy of the branch node d, which is defined as:

g(x_d) Is an S-type function and is defined as:

π (. cndot.) is a fuzzy intuitive index defined as: