CN103400380B - The single camera submarine target three-dimensional track analogy method of fusion image matrix offset - Google Patents

Abstract

The invention discloses the single camera submarine target three-dimensional track analogy method of a kind of fusion image matrix offset, under single camera image-forming condition, based on Bayes tracking framework, in conjunction with underwater scene depth information and camera motion offset vector information, the 3 D motion trace of simulation submarine target.Under Bayes tracking framework, underwater video is carried out target following, output target's center location parameter, dark primary elder generation checking method is utilized to calculate underwater scene depth information, calculate the SURF feature of background dot in consecutive frame simultaneously, carry out characteristic matching, obtain the camera motion offset vector of consecutive frame, last combining target positional information, underwater scene depth information and the 3 D motion trace of camera motion offset vector information simulation submarine target.The present invention can true in single camera video, be reliably achieved the simulation of the 3 D motion trace to submarine target, and operation efficiency is high.

Description

The single camera submarine target three-dimensional track analogy method of fusion image matrix offset

Technical field

The present invention relates to the single camera submarine target 3 D motion trace simulation of a kind of fusion image matrix offset Method, belongs to technical field of machine vision.

Background technology

In recent years, along with developing rapidly of science and technology, DV and the digital photographing of digital technology are used The principle of machine has been applied successfully to Underwater Camera, the manufacturing and designing of underwater camera, investigates in deep-sea science It is widely applied in developing with coastal ocean.Under water in machine vision, the application of higher level needs fixed The position of target in the every two field picture in position, and target following key technology the most therein, last result is The movement locus of energy simulated target.

The existing minority algorithm being capable of the simulation of submarine target three-dimensional track all uses multiple-camera stereopsis Feel technology, and require extremely harsh camera calibration.Cause the hardware complexity of algorithm and computation complexity relatively Height, is difficult to meet the needs of routine application.Additionally, being widely used of mobile camera brings new challenge. Along with the movement of video camera, it is not only the object of motion, and the position array of whole video image all there occurs Skew.Therefore, in this case, common Method for Underwater Target Tracking is the most applicable, but needs The objective course deviation caused due to camera motion skew is compensated.

Based on the problems referred to above, a kind of by obtaining the camera motion offset vector between adjacent video frames, and will It is necessary with the exploitation of the method for simulation submarine target 3 D motion trace as important parameter.

Summary of the invention

Goal of the invention: the technical problem to be solved is to provide a kind of submarine target 3 D motion trace mould Plan method, by combining target positional information, underwater scene depth information and camera motion offset vector letter Breath realizes the simulation of the 3 D motion trace to submarine target.

Summary of the invention: for solving above-mentioned technical problem, the technical solution adopted in the present invention is:

The single camera submarine target three-dimensional track analogy method of fusion image matrix offset, comprises the steps:

Under single camera image-forming condition, realize the tracking to submarine target, output based on Bayesian filter framework Underwater target position coordinate, utilizes dark primary elder generation checking method to calculate underwater scene depth information, calculates adjacent simultaneously In frame, the SURF feature of background dot, carries out images match based on this feature, obtains camera motion offset vector, Last by camera motion offset vector correction underwater target position coordinate, export real submarine target central point Position coordinates, in conjunction with underwater scene depth information, the 3 D motion trace of simulation submarine target.

Wherein, the calculating process of camera motion offset vector based on SURF feature is: calculate in consecutive frame The SURF characteristic point of image background, constructs its characteristic vector, then uses Euclidean distance pair each characteristic point Characteristic vector carries out similarity measurement, obtains distance set, sets threshold value, carries out characteristic matching, finally to phase In adjacent frame, the characteristic point of all couplings is subtracted each other respectively, obtains a distance difference set, then calculates its meansigma methods, Just the motion excursion vector of video camera has been obtained

Beneficial effect: the present invention is a kind of mould carrying out submarine target 3 D motion trace first based on single camera Plan method, the method its in the case of common monocular video, it is possible to true, be reliably achieved submarine target Carry out 3 D motion trace simulation, the submarine target 3 D motion trace analogy method of the present invention, significantly reduce Follow the tracks of the complexity that system hardware is built, and need not loaded down with trivial details camera calibration, the computation complexity of algorithm It is significantly reduced, during therefore this method can be loaded into this underwater video system more widely, Technique Popularizing Property significantly improves.

Accompanying drawing explanation

Fig. 1 is the flow chart of submarine target 3 D motion trace analogy method of the present invention；

Fig. 2 is calculating camera motion offset vector in submarine target 3 D motion trace analogy method of the present invention Flow chart；

Fig. 3 is 9*9 square frame filter template.

Detailed description of the invention

Below in conjunction with specific embodiment, it is further elucidated with the present invention, it should be understood that these embodiments are merely to illustrate this Invention rather than restriction the scope of the present invention, after having read the present invention, those skilled in the art are to this The amendment of the bright various equivalent form of values all falls within the application claims limited range.

As it is shown in figure 1, the single camera submarine target three-dimensional track analogy method of fusion image matrix offset, bag Include following steps:

Under single camera image-forming condition, realize the tracking to submarine target, output based on Bayesian filter framework Underwater target position coordinate, utilize dark primary elder generation checking method calculate underwater scene depth information, i.e. submarine target with And the distance between background and video camera, calculate the SURF feature of background dot in consecutive frame simultaneously, special based on this Levy and carry out images match, obtain camera motion offset vector, by camera motion offset vector correction mesh under water Cursor position coordinate, exports real submarine target center position, in conjunction with underwater scene depth information, simulation The 3 D motion trace of submarine target.

Wherein, the calculating process of camera motion offset vector based on SURF feature mainly includes following three Step: the generation of feature point detection, Feature Descriptor, the coupling of adjacent video frames characteristic point and camera motion are inclined Move the calculating of vector.

Application SURF feature carries out in feature point detection, and SURF algorithm uses Hessian matrix to extract spy Levy a little:

$H=\left[\begin{array}{cc}{L}_{\mathrm{xx}}(x,\mathrm{\σ})& L_{\mathrm{xy}}(x,\mathrm{\σ})\\ L_{\mathrm{xy}}(x,\mathrm{\σ})& L_{\mathrm{yy}}(x,\mathrm{\σ})\end{array}\right]$

L(x,σ)=G(σ)*I(x), $G\left(\mathrm{\σ}\right)={\∂}^{2}g\left(\mathrm{\σ}\right)/\∂x^2$

Wherein, σ is yardstick, and g (σ) is two-dimensional Gaussian function, and L (x, σ) is the volume of G (σ) and integral image Long-pending.

In SURF algorithm, replace second order Gauss filtering, the side of 9*9 by square frame filtering (box filters) approximation Frame wave filter is as it is shown on figure 3, scale factor σ=1.2 of corresponding second order Gauss filtering.On original image, pass through Expand square frame filter size and form the image pyramid of different scale, and use integral image that image convolution is entered Row accelerates, and solves further and obtains Hessian determinant of a matrix:

detH=D_xxD_yy-(0.9D_xy)²

The extreme point that Hessian matrix is detected, 8 consecutive points of each extreme point and unified yardstick thereof with And each 9 points of its upper and lower two yardstick, form the three-dimensional neighborhood of a 3*3*3.By each extreme point with vertical In body field, remaining 26 point compares, only when the value of extreme point is more than all 26 consecutive points, Just using this extreme point as candidate feature point.In metric space, interpolation arithmetic is carried out after obtaining candidate feature point, Obtain stable characteristic point position and place scale-value.

In order to ensure rotational invariance, first obtain characteristic point direction.Structure is with characteristic point as the center of circle, and (s is 6s The yardstick of characteristic point) it is the neighborhood of the radius little wave response of Haar in x and y direction, and give these response values With different Gauss weight coefficients so that the response the closer to characteristic point is contributed the biggest, then by 60 ° of interior x It is added with the y direction little wave response of Haar and forms a local direction vector, travel through whole border circular areas, finally Selecting long vector direction is this feature point principal direction.

Behind selected characteristic point direction, centered by characteristic point, construct the square field of a length of 20, by this window Mouth neighborhood is divided into the subregion of 4*4, each region is calculated to horizontal direction and the Vertical Square of 5*5 sampled point To the little wave response of Haar, be denoted as dx and dy respectively, and give weight with Gauss window function to response value Coefficient.Obtain four-dimensional vector: a V=(∑ d_x,∑d_y,∑|d_x|,∑|d_x|).To each characteristic point 16 sub regions are the formation of the description vectors of 64 dimensions, are carrying out the normalized of vector, are forming characteristic point Description son.

The similarity measurement of two characteristic vectors uses Euclidean distance to calculate:

$D_{\mathrm{ij}}={\left(\underset{k=0}{\overset{N}{\mathrm{\Σ}}}{(X_{\mathrm{ik}}-X_{\mathrm{jk}})}^2\right)}^{1/2}$

In formula: X_ikRepresent the kth element of ith feature point character pair vector, X in previous frame image_jkAfter being The kth element of jth Feature point correspondence characteristic vector in one two field picture, N is characterized the dimension of vector.

For the characteristic vector of the Feature point correspondence of previous frame image, calculate it and all characteristic points of a later frame image In characteristic of correspondence vector set, the Euclidean distance of each characteristic vector, obtains distance set, then adjusts the distance Set carries out sequence from small to large.Set a threshold value, when minimum Eustachian distance and time minimum Eustachian distance When ratio is less than the threshold value set, it is believed that the two characteristic point is coupling.Threshold value is chosen the least, mates logarithm Mesh is the fewest, but more stable.The characteristic point assuming the coupling that t frame and t-1 frame obtain isWithTwo Person is for comprising x, and the three-dimensional coordinate of y, z is vectorial.Finally the characteristic point of all couplings is subtracted each other respectively, Obtain a set, then calculate its meansigma methods, just obtain the offset vector of camera:

${\mathrm{\δ}}_t=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}(p_t^i-p_t^{i-1})$

With the image coordinate system in the first frame for initial coordinate system and as reference, i.e. assume δ₁=(0,0,0), Having been obtained the motion excursion vector of camera by above formula, thus revised the coordinate position of t frame, correction formula is as follows:

$x_t^{\′}=x_t-{\mathrm{\δ}}_t$

Finally combine underwater scene depth information and submarine target center position coordinates just can be in image coordinate system Output can the three-dimensional track of simulated target movement tendency.

Claims (1)

1. a single camera submarine target three-dimensional track analogy method for fusion image matrix offset, its feature exists In: comprise the steps:

Under single camera image-forming condition, realize the tracking to submarine target, output based on Bayesian filter framework Underwater target position coordinate, utilizes dark primary elder generation checking method to calculate underwater scene depth information, calculates adjacent simultaneously In frame, the SURF feature of background dot, carries out images match based on this feature, obtains camera motion offset vector, Last by camera motion offset vector correction underwater target position coordinate, export real submarine target central point Position coordinates, in conjunction with underwater scene depth information, the 3 D motion trace of simulation submarine target；

Wherein, the calculating process of described camera motion offset vector based on SURF feature is: calculate adjacent The SURF characteristic point of image background in frame, constructs its characteristic vector to each characteristic point, then use Euclidean away from Carry out similarity measurement to characteristic vector, obtain distance set, set threshold value, carry out characteristic matching, finally The characteristic point of couplings all in consecutive frame is subtracted each other respectively, obtains a distance difference set, then it is average to calculate it Value, has just obtained the motion excursion vector of video camera.