CN105654514A - Image target tracking method - Google Patents

Abstract

The invention provides an image target tracking method, which comprises the following steps: obtaining a first frame of image, integrating the appearance models of the local image blocks of different scales into a whole image dictionary, and calculating the sparse coefficient of each local image block; collecting the current state of the image as a candidate target, establishing a particle filter and a similarity function, and calculating the estimation position of a candidate target through the similarity function in a current state particle filter frame; and taking the estimation position of the candidate target as a basis to carry out the spare coefficient recalculation of the local image block to finish the final positioning position of the target. The image target tracking method improves a traditional tracking method which carries out the random sampling consistency identification of the target through statistical characteristics, and provides a way that at a target object in a complex background is identified through a distinguishing type model constructed by the spare coefficients of the local image blocks. Through a candidate target similarity function fused by a distinguishing type and a generation type, the stability and the accuracy of the tracking method are improved.

Description

Image target tracking method

Technical Field

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

Background

With the continuous development of sensor technology, computer hardware processing capability and storage technology in recent years, moving target tracking becomes a popular research field in pattern recognition and computer vision, and has wide application in the fields of military, civil use and the like. Visual information is one of the most main information acquired from the outside by human beings through senses, and meanwhile, a video sequence has more effective information than a static image, so that segmenting and tracking a target in the video sequence is the premise and the basis of subsequent research work, and the tasks such as abnormal behavior detection, target identification and the like are all the premise of tracking and segmenting the target.

The research content of the target tracking technology includes two aspects, one is to detect, track, identify and extract the required information of the moving target in the captured video sequence, such as the track of the target and the related motion parameters thereof, such as the speed, the acceleration, the position at a certain moment, and the like. And the other is to estimate and predict the target by using the acquired motion parameters to assist decision making. Therefore, the accurate extraction of the characteristics of the moving target is a precondition for improving the tracking, identification and classification of the target; the tracking accuracy affects the accuracy and difficulty of high-level decision making processes, such as description, understanding, decision making and the like of target behaviors.

However, in reality, a large number of factors affecting the accuracy of extracting the moving target and the appearance features of the moving target exist, for example, the performance of most trackers is affected by the fact that the target is shielded, the target rotates in space, and the target enters and exits the visual field range, so that errors are gradually accumulated, and the final tracking result has large deviation; a crowded scene has a large amount of interference of similar targets, so that the difficulty of detecting and tracking the targets is increased; factors which interfere with the extraction of the edge features of the target, such as the change of the posture of the moving target, can influence the performance of the tracker based on the edge, gradient and other features, so that the tracker cannot detect the area similar to the training sample and influence the tracking result; for example, the jitter and shadow of a camera can cause the gray level and content of two adjacent frames of a video sequence to be greatly changed, and most of the performance of a tracker can be influenced; in addition, conditions such as sudden change of illumination, weak light, low visibility, or contrast change of an indoor scene target and a background can cause changes of target color characteristics, and due to sensitivity of visual characteristics to illumination changes, operation results of some color characteristic-based trackers can be affected. Therefore, an effective solution for dealing with a plurality of factors influencing the performance of the target tracking algorithm in a complex scene is found, the robustness and the accuracy of the algorithm are improved, an efficient and real-time moving target tracking platform is successfully built, and the method has important theoretical value and wide application value for researches such as behavior pattern understanding.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a target tracking method based on a sparse coefficient, and mainly aims at the defect of the identification capability of the existing tracking method for identifying a target by adopting statistical characteristics to carry out random sampling consistency, and provides a mode for identifying a target object in a complex background by using a discrimination model constructed by the sparse coefficient of a local image block. And the stability and the accuracy of the tracking method are improved through the candidate target similarity function fused by the discriminant formula and the generative formula.

In order to achieve the above purpose, the technical invention provides the following technical scheme:

the invention adopts a method for constructing a multi-scale image whole dictionary of a scene image, a static dictionary is established through a target image of a first frame based on different scale blocks, and appearance image representation is realized by combining the image whole dictionary with different scales and local block sparse coefficients.

Furthermore, the invention adopts a discrimination formula method constructed by sparse coefficients of local image blocks. The target representation is realized by the local sparse coefficient through a pool arrangement method, and the single-scale sparse representation is expanded into the sparse coefficient representation under the multi-scale, so that the target appearance representation is perfected, and the robustness of the tracker is enhanced. Considering that the block size selection has a large influence on the tracking performance, the method can solve the problem of image local block selection through a multi-scale block self-adaptive fusion method.

Furthermore, the invention discloses a target state estimation method for establishing a local image sparse histogram, wherein in a production model, the appearance of a target is represented by a corresponding sparse coding histogram, and the similarity between the target template and a candidate template is calculated through the sparse coding histogram.

The invention adopts a framework of a particle filter, calculates the similarity of the candidate target on the basis of tracking the candidate target image and determines the final estimated position of target positioning by combining the particle filter.

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

The invention adopts a target positioning method combining a multi-scale image whole 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 the final positioning position of the candidate target is carried out by combining the particle filter;

furthermore, the invention adopts a target positioning method based on a multi-scale dictionary and particle filtering. And reconstructing a sparse coefficient for each local image block of the candidate target and completing the selection of the optimal position of the tracking target through particle filtering.

Furthermore, the invention adopts a method for effectively identifying the target by constructing a discriminant model through sparse coefficients for tracking the candidate target, the target image in the discriminant model is represented by corresponding sparse codes, and a linear classifier is obtained by learning the sparse codes to identify the background image and the target image.

In conclusion, the target positioning method combining the multi-scale image whole dictionary and the particle filter is adopted for tracking the candidate target, the method not only overcomes the trouble of local block selection before target tracking, but also enables target information and spatial information to be mutually supplemented, and the tracking precision is improved.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:

FIG. 1 is a flowchart of an image target tracking method according to an embodiment of the present invention

Detailed Description

The following detailed description of embodiments of the invention refers to the accompanying drawings. It should be understood that the detailed description and specific examples, while indicating the present invention, are given by way of illustration and explanation only, not limitation. 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 aims at the defect of the identification capability of the existing tracking method for identifying a target by adopting statistical characteristics to carry out random sampling consistency, and provides a mode for identifying a target object in a complex background by using a discriminant model constructed by sparse coefficients of local image blocks. And the stability and the accuracy of the tracking method are improved through the candidate target similarity function fused by the discriminant formula and the generative formula.

In order to achieve the purpose, the invention adopts a method for realizing appearance representation by combining dictionaries with different scales and local block sparse coefficients, and improves the stability and precision of tracking by applying a particle filtering mode, and the adopted technical scheme mainly comprises three parts:

(1) the image appearance model is represented based on sparse coefficients of the multi-scale block:

an image appearance model is established according to the sizes of blocks with different scales, the sparse coefficient of a local image block is solved through the image under different scales, and all target sparsity is collected to express the image appearance. Sampling local image blocks by sliding windows with different sizes in a target area so as to establish an image whole dictionary under different scales $D^s=\left\{d_j^s\right|j=1:n\×K,s=1,2,...,L,r=2\×s+2\}.d_j^s\∈R^d$ Representing the image block of the j-th column, d^sIs the image block dimension and r represents the scale. Then extracting local image blocks from the candidate target area under different scales, and establishing an image whole dictionary D according to the different image blocks^sLocal image blocks of each scale may obtain corresponding sparse coefficients:

$a_j^s=\arg min\left|\right|a_r^s\left|\right|subjectto\left|\right|p_j^s-D^s{a}_j^s\left|\right|<\mathrm{\&epsiv;}$

when a candidate target image block is calculated by the above formula to obtain corresponding sparse coding, the change condition of local images under different scales can be represented by sparse coefficient weighting operation, so that a representation model of the target scene is constructed.

(2) Establishing a particle filter:

the particle filter provides an environment framework for estimating a state change posterior probability density function of the candidate target, and the observation state of the candidate target when the current state is t time is y_s＝{y₁,...,y_r}, then the current state S_tEstimate by maximizing the posterior probability:

s_t＝argmaxp(S_t|y_tr)

where p (S)_t|y_tr) Is the posterior probability that the candidate target is in the given state S_tY of_tThe similarity function of (c).

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

in practical application, similarity calculation is realized by establishing a static expression model through a static dictionary created from a first frame target image based on different scale blocks. And at the time t, the candidate target combines a similarity function in a particle filter frame through a multi-scale dictionary updated in real time to complete the position estimation of the candidate target. Then, on the basis of the estimation result of the first step with the built static dictionary, recalculating the corresponding sparse coefficient for each local image block of the candidate target in the following way:

${\hat{a}}_j^s=\arg \min \left|\right|a_r^s\left|\right|subjectto\left|\right|p_j^s-D^s{a}_j^s\left|\right|<\mathrm{\&epsiv;}$

finally, the final position of the image target is positioned similarly to the first algorithm flow.

In summary, according to the method for tracking the image target provided by the invention, the target state estimation is realized through the appearance similarity and the particle filter tracking framework, a discrimination method constructed through the sparse coefficients of the local image blocks is adopted, the local sparse coefficients are used for realizing target representation through a pool arrangement method, and single-scale sparse representation is expanded into multi-scale sparse coefficient representation, so that the target appearance representation is perfected, and a multi-scale dictionary of the scene image is constructed. Calculating the similarity of candidate targets on the basis of target state estimation, determining final target positioning by combining a particle filter, reconstructing a sparse coefficient for each local image block of the candidate targets, and completing selection of the optimal position of a tracking target through particle filtering. By the target positioning method combining the multi-scale image whole dictionary and the particle filter, the state estimation of the candidate target is analyzed through the multi-scale image whole dictionary integrating the discriminant model and the production model, and the tracking precision is improved by combining the particle filter.

It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.

Claims (5)

1. A method for tracking an image target, comprising the steps of:

(1) acquiring a first frame of image, assembling an image whole dictionary according to appearance models of local image blocks with different scales, and calculating sparse coefficients of the local image blocks;

(2) acquiring the current state of an image as a candidate target, establishing a particle filter and a similarity function, and calculating the estimated position of the candidate target through the similarity function in the current state particle filter frame;

(3) and recalculating the sparse coefficient of the local image block by taking the estimated position of the candidate target as a basis to finish the final positioning position of the target.

2. The method for tracking the image target according to claim 1, wherein the image global dictionary in the step (1) comprises:

sampling local image blocks by sliding windows with different sizes in the appearance model of the first frame image, and establishing an image whole dictionary: $D^s=\left\{d_j^s\right|j=1:n\&times;K,s=1,2,...,L,r=2\&times;s+2\}.$

wherein,representing the image block of the j-th column, d^sIs the image block dimension and r represents the scale. Extracting local image blocks from the candidate target area under different scales, and establishing an image whole dictionary D according to the different image blocks^s。

3. The method for tracking the image target according to claim 1, wherein the sparse coefficients in the step (1) are: $a_j^s=\arg \min \left|\right|a_r^s\left|\right|subjectto\left|\right|p_j^s-D^s{a}_j^s\left|\right|<\mathrm{\&epsiv;}.$

4. the method for tracking image target according to claim 1, wherein the method for calculating the estimated position of the candidate target in step (2) comprises:

the particle filter provides an environmental framework for estimating and diffusing state change posterior probability density functions, and a target observation state y is given until t time_s＝{y₁,...,y_r}；

At the current state S_tEstimate by maximizing the posterior probability: s_t＝argmaxp(S_t|y_tr)；

Wherein p (S)_t|y_tr) As a function of a posterior probability, i.e. given a state S_tLower, y_tThe similarity function of (c).

5. The method for tracking the image target as claimed in claim 1, wherein the method for recalculating the sparse coefficient in the step (3) comprises:

at time t, the candidate target recalculates the corresponding sparse coefficient through each local image block of the estimated position of the target by:

${\hat{a}}_j^s=\arg \min \left|\right|a_r^s\left|\right|subjectto\left|\right|p_j^s-D^s{a}_j^s\left|\right|<\mathrm{\&epsiv;}.$