CN105654514A - Image target tracking method - Google Patents

Abstract

The invention provides a method for image target tracking. By acquiring the first frame of image, the appearance models of local image blocks of different scales are assembled into an overall image dictionary, and the sparse coefficients of local image blocks are calculated; the current state of the collected image is candidate target, establish a particle filter and a similarity function, and calculate the estimated position of the candidate target through the similarity function in the current state particle filter framework; based on the estimated position of the candidate target, recalculate the sparse coefficient of the local image block, Complete the final positioning position of the target. The present invention improves the previous tracking method of random sampling and consistent recognition of targets by using statistical features, and provides a way of recognizing target objects in complex backgrounds through discriminative models constructed by sparse coefficients of local image blocks. And the stability and accuracy of the tracking method are improved through the fusion of discriminative and generative candidate target similarity functions.

Description

A Method of Image Target Tracking

technical field

The invention relates to the technical field of computer image processing, in particular to an image target tracking method in computer image processing.

Background technique

With the continuous development of sensor technology, computer hardware processing capability and storage technology in recent years, moving target tracking has become a popular research field in pattern recognition and computer vision, and has a wide range of applications in military and civilian fields. Visual information is one of the most important information that humans obtain from the outside world through the senses. At the same time, video sequences have more effective information than static images. Therefore, the segmentation and tracking of targets in video sequences is the premise and basis of subsequent research work, such as Abnormal behavior detection, target recognition and other tasks are based on the tracking and segmentation of targets.

The so-called target tracking technology, the research content includes the following two aspects, one is to detect, track, identify and extract the required information for the moving target in the captured video sequence, such as the trajectory of the target and its related motion parameters, such as speed, acceleration , position at a certain moment, etc. The other is to use the acquired motion parameters to estimate and predict the target to assist decision-making. Therefore, accurately extracting the features of moving targets is the prerequisite for improving target tracking, recognition and classification; and the accuracy of tracking affects the accuracy and difficulty of high-level decision-making processes, such as description, understanding and judgment of target behavior, decision-making, etc.

However, in reality, there are a large number of factors that affect the accuracy of moving targets and their shape feature extraction, such as target occlusion, rotation in space, and targets entering and exiting the field of view will affect the performance of most trackers, resulting in gradual accumulation of errors and final tracking. There will be a large deviation in the result; there will be a large number of interferences of similar targets in a crowded scene, which will make the detection and tracking of the target more difficult; many factors that interfere with the extraction of target edge features, such as changes in the posture of moving targets will affect the performance based on edges, gradients, etc. The performance of the feature tracker makes it unable to detect areas similar to the training samples and affects the tracking results; for example, camera shake and shadows will cause grayscale and large changes in the content of two adjacent frames of the video sequence, which will affect most Tracker performance; In addition, such as sudden changes in lighting, weak light, low visibility, or changes in the contrast between the indoor scene target and the background will cause changes in the color characteristics of the target. Due to the sensitivity of visual features to light changes, it will be Affects the results of some trackers based on color features. Therefore, finding an effective solution to many factors that affect the performance of the target tracking algorithm in complex scenes, improving the robustness and accuracy of the algorithm, and successfully building an efficient and real-time moving target tracking platform have important implications for the understanding of behavior patterns. Very important theoretical value and extensive application value.

Contents of the invention

Aiming at the above-mentioned problems existing in the prior art, the present invention provides a target tracking method based on sparse coefficients, mainly aiming at the lack of recognition ability of the existing tracking method for random sampling and consistent identification of targets by using statistical features, and proposes A discriminative model constructed by the sparse coefficients of local image blocks to identify target objects in complex backgrounds. And the stability and accuracy of the tracking method are improved through the fusion of discriminative and generative candidate target similarity functions.

In order to realize the above-mentioned purpose of the invention, the technical invention provides the following technical solutions:

The present invention adopts a method for constructing a multi-scale image overall dictionary of a scene image. Based on blocks of different scales, a static dictionary is created through the target image of the first frame, and the appearance image is realized by combining the overall image dictionary of different scales and the local block sparse coefficient. express.

Further, the present invention adopts a discriminative method constructed by sparse coefficients of local image blocks. The local sparse coefficients are used to realize the target representation through the permutation pooling method, and the single-scale sparse representation is extended to multi-scale sparse coefficient representations, so as to improve the target appearance representation and enhance the robustness of the tracker. Considering that block size selection has a great influence on tracking performance, this method can solve the problem of image local block selection through multi-scale block adaptive fusion method.

Further, the present invention establishes a method for estimating the target state of the local image sparse histogram. In the production model, the appearance of the target is represented by the corresponding sparse-coded histogram, and the similarity between the target template and the candidate template is calculated by the sparse-coded histogram .

The present invention adopts the frame of the particle filter, calculates the similarity degree of the candidate target on the basis of tracking the image of the candidate target, and combines the particle filter to determine the estimated position of the final target location.

Further, the particle filter provides an environmental framework for estimating and diffusing the posterior probability density function of state changes.

The present invention adopts a target positioning method combining a multi-scale image overall dictionary and a particle filter. The state estimation of a candidate target is analyzed through a multi-scale dictionary integrating a discriminant model and a production model, and combined with a particle filter to estimate the candidate target. The final positioning position of the target;

Further, the present invention adopts a method of target location based on multi-scale dictionary and particle filter. The sparse coefficients are reconstructed for each local image block of the candidate target, and the selection of the optimal position of the tracking target is completed through particle filtering.

Further, the present invention adopts an effective target recognition method for constructing a discriminative model through sparse coefficients to track candidate targets. In the discriminative model, target images are represented by corresponding sparse codes. Through the learning of sparse codes, Obtain a linear classifier to identify background and target images.

To sum up, the present invention uses a target positioning method combining multi-scale image overall dictionary and particle filter to track candidate targets. This method not only overcomes the trouble of local block selection before target tracking, but also makes target information It complements the spatial information and improves the tracking accuracy.

Description of drawings

The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the description, together with the following specific embodiments, are used to explain the present invention, but do not constitute a limitation to the present invention. In the attached picture:

Fig. 1 is the flowchart of image target tracking method in the embodiment of the present invention

detailed description

Specific embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described here are only used to illustrate and explain the present invention, and are not intended to limit the present invention. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the invention. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

A target tracking method based on sparse representation, mainly aimed at the lack of recognition ability of the existing tracking methods that use statistical features to randomly sample and consistently identify targets, and propose a method constructed by sparse coefficients of local image blocks Discriminative models to identify the way to target objects in complex backgrounds. And the stability and accuracy of the tracking method are improved through the fusion of discriminative and generative candidate target similarity functions.

In order to achieve the above purpose, this technical invention adopts a method of combining dictionaries of different scales and local block sparse coefficients to realize the appearance representation, and uses particle filtering to improve the stability and accuracy of tracking. The technical solutions adopted mainly include three part:

(1) Image appearance model represented by sparse coefficients based on multi-scale blocks:

The image appearance model is established according to the block size of different scales, the sparse coefficient of the local image block is calculated through the image at different scales, and all the target sparseness is collected to represent the image appearance. In the target area, local image blocks are sampled by sliding windows of different sizes, so as to establish an overall image dictionary at different scales ${\mathrm{D.}}^{\mathrm{the\; s}}=\left\{d_j^{\mathrm{the\; s}}\right|j=1:\mathrm{no}\×K,\mathrm{the\; s}=1,2,...,L,r=2\×\mathrm{the\; s}+2\}.d_j^{\mathrm{the\; s}}\∈R^d$ Indicates the image block in the jth column, d ^s is the dimension of the image block, and r represents the scale. Then extract local image blocks from the candidate target area at different scales, and establish an overall image dictionary D ^s according to different image blocks. The corresponding sparse coefficients can be obtained for each scale of local image blocks:

${\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; b</span>}\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}$

When a candidate target image block is calculated by the above formula and the corresponding sparse coding is obtained, the change of the local image at different scales can be represented by the sparse coefficient weighting operation to construct the representation model of the target scene.

(2) Establishment of particle filter:

The particle filter provides an environmental framework for estimating the posterior probability density function of the state change of the candidate target. The current state is the candidate target observation state at time t is y _s ={y ₁ ,...,y _r }, then the current state S _t is estimated by maximizing the posterior probability:

s _t =argmaxp(S _t |y _tr )

Here p(S _t |y _tr ) is the posterior probability, which is the similarity function of candidate target y _t in a given state S _t .

(3) Track the target according to the multi-scale dictionary:

In practical applications, based on blocks of different scales, a static representation model is established through a static dictionary created from the first frame of the target image to achieve similarity calculation. At time t, the candidate targets are estimated through the multi-scale dictionary updated in real time, combined with the similarity function in the particle filter framework to complete the candidate target position estimation. Then, on the basis of the estimated results of the first step with the static dictionary constructed, the corresponding sparse coefficients are recalculated for each local image block of the candidate target in the following way:

${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; a</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; b</span>}\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}$

Finally, similar to the first algorithm process, the final location of the image target is completed.

In summary, the present invention provides a method for image target tracking, which realizes target state estimation through appearance similarity and particle filter tracking framework, adopts a discriminative method constructed by sparse coefficients of local image blocks, and The local sparse coefficient realizes the target representation through the permutation pooling method, and expands the single-scale sparse representation into a multi-scale sparse coefficient representation to improve the target appearance representation and build a multi-scale dictionary of the scene image. On the basis of target state estimation, the similarity of candidate targets is calculated and combined with the particle filter to determine the final target location, the sparse coefficients are reconstructed for each local image block of the candidate target, and the optimal position of the tracking target is selected through the particle filter. Through the target localization method combining the multi-scale image dictionary and the particle filter, the state estimation of the candidate target is analyzed through a multi-scale image dictionary integrating the discriminant model and the production model, and the particle filter is combined to improve the tracking accuracy.

Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these modifications and variations.

Claims (5)

1. A method for image target tracking, comprising the following steps: (1) Obtain the first frame of image, assemble the overall image dictionary according to the appearance models of local image blocks of different scales, and calculate the sparse coefficient of the local image blocks; (2) Collect the current state of the image as a candidate target, establish a particle filter and a similarity function, and calculate the estimated position of the candidate target through the similarity function in the current state particle filter framework; (3) Based on the estimated position of the candidate target, recalculate the sparse coefficient of the local image block to complete the final positioning position of the target. 2. the method for a kind of image object tracking according to claim 1, is characterized in that, the image whole dictionary in the described step (1) comprises: In the appearance model of the first frame image, the local image block is sampled by sliding windows of different sizes, and an overall image dictionary is established: ${\mathrm{D.}}^{\mathrm{the\; s}}=\left\{d_j^{\mathrm{the\; s}}\right|j=1:\mathrm{no}\&times;K,\mathrm{the\; s}=1,2,...,L,r=2\&times;\mathrm{the\; s}+2\}.$ in, Indicates the image block in the jth column, d ^s is the dimension of the image block, and r represents the scale. Extract local image blocks from candidate target areas at different scales, and build an overall image dictionary D ^s according to different image blocks. 3. the method for a kind of image target tracking according to claim 1, is characterized in that, the sparse coefficient in the described step (1) is: $a_j^{\mathrm{the\; s}}=\arg \min \left|\right|a_r^{\mathrm{the\; s}}\left|\right|\mathrm{the\; s}ubjectto\left|\right|p_j^{\mathrm{the\; s}}-{\mathrm{D.}}^{\mathrm{the\; s}}{a}_j^{\mathrm{the\; s}}\left|\right|<\mathrm{\&epsiv;}.$ 4. the method for a kind of image target tracking according to claim 1, is characterized in that, the method for calculating the estimated position of candidate target in the described step (2), comprises: The particle filter provides an environmental framework for estimating and diffusing the posterior probability density function of state changes, given the target observed state y _s ={y ₁ ,...,y _r } up to time t; Estimated by maximizing the posterior probability with the current state S _t : s _t =argmaxp(S _t |y _tr ); Where p(S _t |y _tr ) is the posterior probability function, that is, the similarity function of y _t under a given state S _t . 5. the method for a kind of image target tracking according to claim 1, is characterized in that, the method for sparse coefficient recalculation in described step (3), comprises: At time t, the candidate target passes through each local image block of the estimated position of the target to recalculate the corresponding sparse coefficient in the following way: ${\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; a</span>}}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; b</span>}\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; .</span>$