CN110276784B - Correlation filtering moving target tracking method based on memory mechanism and convolution characteristics - Google Patents

Abstract

The invention provides a related filtering moving target tracking method based on a memory mechanism and convolution characteristics, and belongs to the technical field of computer vision. The method utilizes a pre-trained deep convolutional neural network to extract the convolutional characteristic of a target, is inspired by a human brain memory mechanism in human visual information processing cognitive behaviors, and integrates the memory mechanism into the detection, training and updating processes of a classifier of a relevant filtering method. The memory mechanism consists of three parts, namely response diagram decision, adaptive peak detection and adaptive fusion coefficient. The method has stronger robustness, and can still continuously and stably realize target tracking under the conditions of violent deformation, reappearance or shielding after temporary disappearance and the like of the target. Meanwhile, the method has higher target tracking speed, reduces the complexity and reduces the operation amount.

Description

Correlation filtering moving target tracking method based on memory mechanism and convolution features

technical field

The invention relates to a tracking method for moving objects in an image sequence, in particular to a related filtering moving object tracking method based on a memory mechanism and convolution features, and belongs to the technical field of computer vision.

Background technique

Moving target tracking technology is an important research direction of computer vision science, which is widely used in security monitoring, human-machine interface, medical diagnosis and other fields. At present, the main problem of moving target tracking technology is that it is difficult to overcome the influence of complex interference factors such as changes in background lighting conditions, target occlusion, shape changes, size changes, and rapid motion, resulting in a decrease in tracking accuracy.

The discriminative tracking method is an important moving target tracking method, which includes: Multiple Instance Learning (MIL) tracking method, Tracking-Learning-Detection (TLD) tracking method, kernel-based structure Output (Structured output tracking withkernel, Struck) tracking method, etc. The principle of this kind of method is: first, take the target as a positive sample and the background as a negative sample, and train a classifier; then, use this classifier to detect the search area, and regard the point with the largest responsiveness as the center of the target, to track. Typically, such methods train the classifier by sparse sampling, that is, taking several equal-sized windows around the target as samples. However, when the number of samples increases, the amount of computation increases, which reduces the real-time performance of the tracking method.

The correlation filtering tracking method, by constructing the cyclic matrix of the samples, solves the problems of insufficient training samples and large amount of computation of the discriminative tracking method to a certain extent. For example, the KCF algorithm proposed by Henriques et al. (Henriques J F, Rui C, Martins P, et al. "High-Speed Tracking with Kernelized Correlation Filters". IEEE Transactions on Pattern Analysis & Machine Intelligence, 2014, 37(3): 583-596), According to the characteristic that the circulant matrix becomes a diagonal matrix after Fourier transform, after shifting and circulating a single sample, the classifier can be quickly detected and trained in the Fourier domain. And through the kernel-based ridge regression operation, the process of correlation filtering is realized. The algorithm not only has high real-time performance, but also achieves accurate tracking of moving targets under nonlinear conditions.

In recent years, research results in the field of deep learning have begun to be combined with correlation filter tracking methods. For example, HCF algorithm (Ma C, Huang J B, Yang X, et al. Hierarchical Convolutional Features for Visual Tracking [C]//IEEE International Conference on Computer Vision. IEEE Computer Society, 2015: 3074-3082.) under the framework of KCF algorithm , replacing HOG features with hierarchical convolutional features. According to the characteristics that high-level features contain more semantic information, and low-level features contain more local information such as textures and contours, first use the top-level features to determine the approximate location of the target, and then gradually and accurately locate the target downward. Compared with the traditional hand-extracted features, Has higher robustness.

Although the correlation filtering algorithm using convolution features has the above advantages, it also has certain limitations: First, the classifier extracts two convolution features during detection and training, which is very computationally expensive; Rate update target templates and classifiers, resulting in poor ability to adapt to drastic changes in the target. Therefore, when the target has sudden changes in shape, severe occlusion, and reappears after a short disappearance, its tracking accuracy will be significantly reduced, or even the target will be lost; and it is difficult to meet the real-time requirements.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the problem of accurate and high-speed tracking of the target under the interference conditions such as the sudden change of the posture and shape of the target, the reappearance after a short disappearance, and the occlusion, and proposes a related filtering moving target based on a memory mechanism and convolution features. The tracking method integrates the memory mechanism of the human brain into the classifier detection, training and update process of the relevant filtering algorithm, which can achieve moving target tracking with high accuracy, strong robustness and fast calculation speed in complex application scenarios.

The method of the present invention utilizes a pre-trained deep convolutional neural network to extract the convolution features of the target. Inspired by the human brain memory mechanism in the cognitive behavior of human visual information processing, the memory mechanism is integrated into the detection, training and update process of the classifier of the correlation filtering method. Among them, the memory mechanism consists of three parts: response map decision, adaptive peak detection and adaptive fusion coefficient. The process of integrating the memory mechanism with the detection, training and updating of the classifier is described as follows:

(1) Classifier detection based on response map decision: After extracting the convolution feature of the candidate region, all classifiers in the memory space are convolved with it to obtain their respective response maps, and select the response map with the largest peak value to target the target. to locate.

(2) Classifier training based on adaptive peak detection: After the target is located, the size relationship and positional relationship between the main peak and the second-highest interference peak in the response graph are synthesized, and the change of the target is analyzed. If the degree of interference is greater than the threshold, the convolutional features of the target are extracted again to train a new classifier. If the degree of interference is not greater than the threshold, the classifier will not be trained and updated.

(3) Classifier update based on adaptive fusion coefficient: After training a new classifier, the fusion coefficient is adaptively calculated according to the result of peak detection. The more severe the interference, the larger the fusion coefficient.

Through the above methods, the organic integration of memory mechanism and tracking method is realized.

The concrete realization process of the method of the present invention is as follows:

A related filtering moving target tracking method based on memory mechanism and convolution features, comprising the following steps:

Step 1: Initialize the memory space.

Let the capacity of the memory space be m, in the 1st to mth frames, the memory space is filled first, and the memory mechanism is not executed for the time being. After completing the training of the classifier in the ith frame, the parameters of the classifier are stored in the memory space as the ith classifier w[i], i∈{1,...,m} in the memory space. Except for initializing the memory space, the steps of the method of the present invention are exactly the same as those of the general correlation filter tracking method. When the memory space is filled, the memory mechanism starts to execute in subsequent frames.

Step 2: Perform classifier detection based on response graph decision.

Step 2.1: Extract the convolutional features of the candidate region of the current frame.

Read the t-th frame image, t>m, and select the candidate area according to the target center position determined in the previous frame. With the help of pre-trained convolutional neural network, the convolutional features of the tracking window are extracted. After inputting the subsequent region images into the convolutional neural network, the output of the L layer among the 19 convolutional layers is selected as the convolutional feature x _t . The feature of the candidate region at the lth layer at time t is denoted as _xt [l], l∈L.

After the convolution feature x _t is extracted, a circulant matrix is constructed with x _t as the generator matrix, so as to obtain the detection sample C(x _t ).

Step 2.2: Detect all classifiers in the memory space.

Let w _t-1 [i,l] denote the parameters of the i-th classifier learned before the t-th frame in the memory space corresponding to the l-th layer feature, i∈{1,...,m}, l∈L . The response map is obtained by convolving the detection sample C(x _t ) with the classifier, and the position of the maximum response value on the response map is regarded as the target position.

From the properties of the circulant matrix, it can be known that the convolution of any matrix and the circulant matrix in the time domain can be expressed as the dot product of the generator matrix of the circulant matrix in the frequency domain. The response f _t [i, l] of each layer feature is added according to the fixed weight, and the response map f _t [i] of the i-th classifier in the memory space at the t-th frame is obtained:

表示快速反傅里叶变换(IFFT)操作，⊙为点乘运算符，大写字母代表变量的傅里叶变换形式，γ是融合权重。X_t[l]代表第t帧时第l层特征的傅里叶变换形式。in,

Indicates the Inverse Fast Fourier Transform (IFFT) operation, ⊙ is the dot product operator, capital letters represent the Fourier transform form of the variable, and γ is the fusion weight. X _t [l] represents the Fourier transform form of the l-th layer feature at the t-th frame.

Make all the classifiers in the memory space perform convolution operation with the cyclic samples to obtain m response maps, take the response map with the largest response peak to infer the target position, and make the classifier corresponding to the response map perform subsequent training and update:

In the formula, π is the index of the classifier corresponding to the maximum peak response map in the memory space.

Step 3: Perform classifier training based on adaptive peak detection.

Step 3.1: Adaptive peak detection.

At the same time, the relationship between the position and peak size of the main peak and the interference peak on the response graph was calculated and compared, and the secondary peak other than the main peak was selected as the interference peak. When the interference peak is far from the main peak, even if the interference peak is high, it is considered that the target is not blocked; when the interference peak appears near the main peak, even if the interference peak is not high, it is judged that the target is blocked. Using the peak interference degree to judge the target state, the formula is as follows:

是干扰峰相对于主峰的位置向量，

为构造的抛物面。如干扰峰高于此曲面，则认为目标发生了剧烈变化。ρ值为干扰峰超过曲面的距离与整个干扰峰高度的比值，若峰值干扰度ρ＝0，则跳过以下所有步骤，不再进行分类器的训练与更新，直接进入下一帧；当峰值干扰度ρ>0时，执行以下步骤：Among them, the response graph redraws the coordinate system with the center of the main peak as the origin, H is the peak value of the main peak on the response graph, h is the peak value of the interference peak, M is the distance from the main peak in the direction of the interference peak to the edge of the response graph,

is the position vector of the interference peak relative to the main peak,

for the constructed paraboloid. If the interference peak is higher than this surface, the target is considered to have changed drastically. The ρ value is the ratio of the distance between the interference peak exceeding the surface and the height of the entire interference peak. If the peak interference degree ρ=0, skip all the following steps, no longer train and update the classifier, and directly enter the next frame; When the degree of interference ρ>0, perform the following steps:

Step 3.2: Extract the convolutional features of the target region of the current frame.

According to the result of positioning the current frame in step 2, take the center of the target as the center, expand the target area with the same size as the subsequent area, input the target area into the convolutional neural network, and extract its convolutional feature x _t '.

Step 3.3: Classifier training.

The peak interference degree ρ>0 indicates that the classifier corresponding to the peak maximum response map selected in step 3.1 has a poor matching degree with the target, and a new classifier _wt ' needs to be trained to adapt to the change of the target.

The principle of training the classifier is the same as the general correlation filtering method. The classifier parameter w _t '[l] corresponding to the feature of the lth layer is trained by minimizing the following formula:

Among them, x' _t [l] is the feature extracted at the new position during training, * is the convolution operator, λ is the l2 regularization parameter; y is the training target label function, which is the same size as the classifier. A 2D Gaussian function with the peak at the center.

The closed-form solution to this minimization problem is:

表示目标标签函数的傅里叶变换形式。in,

Represents the Fourier transform form of the target label function.

Step 4: Perform classifier update based on adaptive fusion coefficients.

After the new classifier parameters w _t ' are trained, the classifier in the memory space is updated. The classifier w _t-1 [π] is weighted and fused with w _t ', and the rest of the classifier parameters remain unchanged, and the formula is as follows:

Among them, λ is the fusion coefficient of the classifier in the current frame, which is obtained adaptively using the Sigmoid function:

Among them, λ increases monotonically with respect to ρ, so that the more severe the change of the target, the faster the update rate of the classifier; e is the natural logarithmic symbol.

beneficial effect

Compared with the existing moving target tracking method, the method of the present invention has the following advantages:

(1) Strong robustness. The method of the invention has strong robustness, and by incorporating the memory mechanism of the human brain into the relevant filtering algorithm, the algorithm can remember the state that the target has appeared in during tracking. On the one hand, the response graph decision is used to select the most suitable classifier from the memory space for detection. On the other hand, the classifier is trained using adaptive peak detection, and the convolution features of the target are re-extracted to retrain the classifier only when the target changes drastically. At the same time, the fusion coefficient is adaptively calculated according to the results of peak detection to classify Therefore, under the conditions of violent deformation of the target, reappearance after a short disappearance or occlusion, etc., the target tracking can still be achieved continuously and stably.

(2) The tracking speed is fast. The method of the invention has higher target tracking speed. On the one hand, under the framework of correlation filtering, the training samples of the classifier are constructed by cyclic offset. At the same time, the problem is transformed to the frequency domain solution based on the characteristics of the circulant matrix, which avoids the matrix inversion process, thus greatly reducing the complexity of the algorithm. On the other hand, the classifier parameters of the target in different states are stored in the memory space. When a similar state appears again, the classifier is directly selected and called according to the response value, and there is no need to extract the CNN features of the target area for retraining, thus reducing the amount of computation by nearly half.

Description of drawings

Fig. 1 is the principle flow chart of the method of the present invention;

Fig. 2 is the principle schematic diagram of the classifier detection step based on the response graph decision in the method of the present invention;

Fig. 3 is the principle schematic diagram of the classifier training step based on adaptive peak detection in the method of the present invention;

4 is a schematic diagram of the principle of updating steps of a classifier based on adaptive fusion coefficients in the method of the present invention;

Fig. 5 is the concrete flow chart of the method of the present invention;

Fig. 6 is the tracking result comparison of the inventive method and conventional HCF method;

Fig. 7 is the tracking precision curve of the inventive method and conventional HCF method;

FIG. 8 is a comparison of the tracking indicators between the method of the present invention and the conventional HCF method.

Detailed ways

The method of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

Example

A related filtering moving target tracking method based on memory mechanism and convolution features, the implementation process is shown in Figure 2, including the following steps:

Step 1: Initialize the memory space.

Let the capacity m of the memory space be 4. In frames 1 to 4, except for initializing the memory space, the method of the present invention is exactly the same as the general correlation filter tracking method. After the training of the classifier is completed in each frame, the parameters of the classifier are stored in the memory space as the ith classifier in the memory space. At the end of frame 4, the memory space is filled, and the memory mechanism begins to execute in subsequent frames.

Step 2: Classifier detection based on response graph decision.

Step 2.1: Extract the convolutional features of the candidate region of the current frame.

Read the t-th frame image, and select the candidate area according to the target center position determined in the previous frame. The method of the invention extracts the convolution feature of the tracking window by using the trained VGG-19 convolutional neural network. After inputting the subsequent region images into the convolutional network, the outputs of Conv3-4, Conv4-4 and Conv5-4 in the 19 convolutional layers are selected as the convolutional features, that is, L={Conv3-4, Conv4-4, Conv5 -4}. The feature of the selected region at the lth layer at time t is represented as _xt [l], l∈L.

After the convolution feature x _t is extracted, a circulant matrix is constructed with x _t as the generator matrix, so as to obtain the detection sample C(x _t ).

Step 2.2: Detection of all classifiers in memory space.

Let w _t-1 [i,l] denote the parameters of the i-th classifier learned before the t-th frame in the memory space corresponding to the l-th layer feature, i∈{1,2,3,4}, l∈L . The response map can be obtained by convolving the detection sample C(x _t ) with the classifier, and the position of the maximum response value on the response map is regarded as the target position.

From the properties of the circulant matrix, it can be known that the convolution of any matrix and the circulant matrix in the time domain can be expressed as the dot product of the generator matrix of the circulant matrix in the frequency domain. The response f _t [i, l] of each layer of features is added according to a fixed weight, and the response map f _t [i] of the i-th classifier in the memory space at the t-th frame is obtained:

表示快速反傅里叶变换(IFFT)操作，⊙为点乘运算符，大写字母代表变量的傅里叶变换形式，γ是融合权重，设为γ＝{0.25,0.5,1}。in,

represents the Inverse Fast Fourier Transform (IFFT) operation, ⊙ is the dot product operator, capital letters represent the Fourier transform form of the variable, and γ is the fusion weight, set to γ={0.25, 0.5, 1}.

All the classifiers in the memory space are convolved with the cyclic samples to obtain m response maps, and the response map with the largest response peak is taken to infer the target position. And make the classifier corresponding to the response map perform subsequent training and update.

where π is the index of the classifier corresponding to the maximum peak response graph in the memory space.

Step 3: Classifier training based on adaptive peak detection.

Step 3.1: Adaptive Peak Detection

The core idea of adaptive peak detection is to calculate and compare the relationship between the position and peak size of the main peak and the interference peak on the response graph at the same time. The secondary peaks other than the primary peak were selected as interference peaks. When the interference peak is far from the main peak, even if the interference peak is high, it is considered that the target is not blocked; when the interference peak appears near the main peak, even if the interference peak is not high, it is judged that the target is blocked. The target state is judged by the peak interference degree, and the calculation formula is as follows:

是干扰峰相对于主峰的位置向量，

为构造的抛物面。如果干扰峰高于此曲面，则认为目标发生了剧烈变化。ρ的值即为干扰峰超过曲面的距离与整个干扰峰高度的比值。若峰值干扰度ρ＝0，则跳过以下步骤，不再进行分类器的训练与更新，直接进入下一帧。In the formula, the response graph re-defines the coordinate system with the center of the main peak as the origin, H is the peak value of the main peak on the response graph, h is the peak value of the interference peak, M is the distance from the main peak in the direction of the interference peak to the edge of the response graph,

is the position vector of the interference peak relative to the main peak,

for the constructed paraboloid. If the interference peak is above this surface, the target is considered to have changed drastically. The value of ρ is the ratio of the distance of the interference peak beyond the surface to the height of the entire interference peak. If the peak interference degree ρ = 0, the following steps are skipped, the classifier training and updating are no longer performed, and the next frame is directly entered.

When the peak interference level ρ>0, the following steps are performed.

Step 3.2: Extract the convolutional features of the target region of the current frame.

According to the positioning result of the current frame in step 2, take the center of the target as the center, and expand the target area with the same size as the subsequent area. The target region is input into the VGG-19 network, and its convolutional features x _t ' are extracted.

Step 3.3: Classifier training.

The peak interference degree ρ>0 means that the classifier corresponding to the peak maximum response map selected in step 3.1 has a poor matching degree with the target, and a new classifier _wt ' needs to be trained to adapt to the change of the target.

The principle of training the classifier is the same as the general correlation filtering method. The classifier parameter w _t '[l] corresponding to the feature of the lth layer is trained by minimizing the following formula:

where * is the convolution operator, λ is the l2 regularization parameter, y is the training target label function, a two-dimensional Gaussian function with the same size as the classifier, and the peak is at the center.

The closed-form solution to this minimization problem is:

Step 4: Classifier update based on adaptive fusion coefficients.

After the new classifier parameters w _t ' are trained, the classifier in the memory space is updated. The classifier w _t-1 [π] is weighted and fused with w _t ', and the rest of the classifier parameters remain unchanged, which is described by the formula as:

Among them, λ is the fusion coefficient of the classifier in the current frame, which is obtained adaptively using the Sigmoid function:

λ increases monotonically with respect to ρ, so that the more drastic the target changes, the faster the classifier updates.

The simulation effect of the present invention is illustrated by the following simulation experiments:

1. Simulation conditions:

The present invention uses the MATLAB 2017b platform on the PC of Intel(R) Core(TM) i7-7700HQ CPU 2.80GHz, RAM 8.00G, GTX1050GPU to complete the simulation experiment on the video sequences in the Visual Tracker Benchmark video test set.

2. Simulation results:

Fig. 3 is a graph of the tracking result of a video sequence with obvious occlusion of the target, which are the 330th, 371st, 390th and 410th frames respectively. The rectangular boxes in the figure represent the tracking results of the conventional method and the method of the present invention. It can be seen from FIG. 3 that the present invention can accurately track the moving target in the process of reappearing after obvious occlusion.

FIG. 4 is a comparison diagram of the tracking accuracy curve between the method of the present invention and the conventional HCF algorithm. The abscissa of the tracking accuracy curve refers to the Euclidean distance between the target center of the simulation tracking result and the real center marked in the groundtruth, and the ordinate refers to the ratio of the number of frames whose Euclidean distance is less than a certain threshold to the length of the entire test video sequence. Figure 5 is a graph of tracking accuracy versus tracking speed (FPS: frames per second) at a distance threshold of 20 pixels. According to the evaluation statistics, for the Lemming sequence, the probability that the distance between the tracking result of the conventional HCF algorithm and the method of the present invention and the actual position of the target within 20 pixels is 0.6820 and 0.8920, respectively, and the tracking accuracy is improved by 30.8%. When the CNN operation is completed on the GPU, the speed of the conventional HCF algorithm and the algorithm proposed by the present invention are 4.4751fps and 5.1678fps, an increase of 15.5%; when the CNN operation is completed on the CPU, the speed of the two algorithms is 1.1653 fps and 2.1363fps, an improvement of 83.3%.

Claims (3)

1. A correlation filtering moving target tracking method based on a memory mechanism and convolution characteristics is characterized by comprising the following steps:

firstly, initializing a memory space, and the method comprises the following steps:

setting the capacity of a memory space as m, filling the memory space when the frames 1 to m are processed, not executing a memory mechanism for the moment, storing parameters of a classifier into the memory space after the training of the classifier is completed in the ith frame, using the parameters as an ith classifier w [ i ], wherein i belongs to { 1., m } in the memory space, and starting executing the memory mechanism in a subsequent frame when the memory space is filled;

secondly, extracting the convolution characteristics of the target by utilizing a pre-trained deep convolutional neural network, and integrating a memory mechanism into the detection, training and updating fusion process of a classifier of a related filtering method, wherein the memory mechanism consists of three parts, namely response diagram decision, adaptive peak detection and adaptive fusion coefficients; specifically, a memory mechanism is integrated into a detection, training and updating fusion process of a classifier of a relevant filtering method, and the method comprises the following steps:

the classifier detection based on the response graph decision comprises the following steps: after extracting the convolution characteristics of the candidate region, carrying out convolution operation on all classifiers in the memory space with the candidate region to obtain respective response graphs, and selecting the response graph with the maximum peak value to position the target;

the classifier training based on the self-adaptive peak detection comprises the following steps:

step A: self-adaptive peak detection;

meanwhile, the position and peak value size relation of the main peak and the interference peak on the response graph is calculated and compared, and the secondary peak except the main peak is selected as the interference peak; when the interference peak is far away from the main peak, even if the interference peak is high, the target is considered not to be shielded, and when the interference peak is close to the main peak, the target is judged to be shielded even if the interference peak is not high;

and judging the target state by using the peak interference degree, wherein the formula is as follows:

wherein, the response diagram uses the center of the main peak as the origin to define the coordinate system again, H is the peak value of the main peak on the response diagram, H is the peak value of the interference peak, M is the distance from the main peak to the edge of the response diagram in the direction of the interference peak,

is the position vector of the interference peak relative to the main peak,

for a constructed paraboloid, if the interference peak is higher than the curved surface, the target is considered to be changed violently; the rho value is the ratio of the distance of the interference peak exceeding the curved surface to the height of the whole interference peak, if the peak value interference degree rho is 0, the subsequent steps are skipped, the training and updating of the classifier are not carried out, and the next frame is directly entered; when peak interference degree rho>0, executing the following steps:

and B: extracting convolution characteristics of a current frame target area;

according to the positioning result of the current frame in the classifier detection process, a target center is used as a center, a target area with the same size as a subsequent area is obtained through expansion, the target area is input into a convolutional neural network, and the convolutional characteristic x of the target area is extracted_t'；

And C: training a classifier;

peak interference degree rho>0, indicating that the degree of matching between the classifier corresponding to the peak maximum response graph selected in the step A and the target is poor, and a new classifier w needs to be trained_t' to accommodate changes in goals;

training classifier parameters w corresponding to the l-th layer features by minimizing the following formula_t'[l]：

Wherein, x'_t[l]Features extracted at new positions during training are convolution operators, and lambda is a l2 regularization parameter; y is a trained target label function and is a two-dimensional Gaussian function with the same size as the classifier, and the peak value is positioned at the center;

the closed-form solution to this minimization problem is:

wherein,

a Fourier transform form representing a target tag function;

updating a classifier based on a self-adaptive fusion coefficient, wherein the method comprises the following steps: after a new classifier is trained, a fusion coefficient is calculated in a self-adaptive mode according to a peak detection result, and the more severe the interference is, the larger the fusion coefficient is.

2. The method for tracking the moving object based on the correlation filtering of the memory mechanism and the convolution characteristic as claimed in claim 1, wherein the classifier detection based on the response graph decision is as follows:

step 2.1: extracting convolution characteristics of current frame candidate region

Setting the memory space capacity as m; reading the t-th frame image, t>m, selecting a candidate area according to the target center position determined by the previous frame; extracting the convolution characteristic of the tracking window by means of a pre-trained convolution neural network; after the subsequent region image is input into the convolutional neural network, the output of L layers in 19 convolutional layers is selected as the convolutional characteristic x_tAnd the characteristic of the candidate region at the ith layer at the moment t is represented as x_t[l]，l∈L；

Extracting convolution features x_tThen, with x_tConstructing a circulant matrix for the generated matrix to obtain a test sample C (x)_t)；

Step 2.2: detecting all classifiers in a memory space

Let w_t-1[i,l]Parameters representing the ith classifier learned before the tth frame in memory space corresponding to the ith layer features, i ∈ { 1., m }, L ∈ L, are used to detect the sample C (x)_t) Convolving with a classifier to obtain a response graph, and regarding the position of the maximum response value on the response graph as a target position;

the response f of each layer characteristic_t[i,l]Adding according to fixed weight to obtain the response image f of the ith classifier in the memory space at the t frame_t[i]：

Wherein,

denotes an Inverse Fast Fourier Transform (IFFT) operation, which is a dot product operator, capital letters denote Fourier transform forms of variables, and γ is a fusion weight; x_t[l]A Fourier transform form representing the characteristics of the l < th > layer at the t < th > frame;

performing convolution operation on all classifiers in the memory space and the cyclic sample to obtain m response graphs, estimating a target position by taking the response graph with the maximum response peak value, and performing subsequent training and updating on the classifier corresponding to the response graph:

in the formula, pi is the index of the classifier corresponding to the maximum peak response graph in the memory space.

3. The method for tracking the moving object based on the correlation filtering of the memory mechanism and the convolution characteristic as claimed in claim 1, wherein the classifier updating method based on the adaptive fusion coefficient is as follows:

new classifier parametersNumber w_tAfter training, updating the classifier in the memory space; classifier w_t-1[π]And w_t' carry out weighted fusion, and the parameters of the rest classifiers are unchanged, and the formula is as follows:

wherein λ is a fusion coefficient of the classifier at the current frame, and is obtained by using a Sigmoid function in a self-adaptive manner:

wherein λ monotonically increases with respect to ρ such that the more drastic the change in the target, the faster the rate of classifier update; e is a natural log symbol.

Images