Target Tracking Algorithm Based On The Joint Model

In recent years,with the development of Internet and computer technology,computer vision has also made great strides.Video-based object tracking is a very important part of computer vision.It is widely used in military,security,transportation and many other fields.Although researchers have done a great deal of research in the past few decades,there have been many good achievements in the tracking of targets for simple scenes.However,in actual scenes,we are confronted with the target occlusion,the target self-deformation,and the rapid movement of the target.Complex backgrounds and many other challenges.These challenges will lead to our algorithm in the process of tracking the target will appear unstable or even wrong.Therefore,it is still a challenging task to devise a strategy that can meet a variety of challenges in a real world and still be able to track targets steadily and efficiently.In view of the above problems,in this paper,we improve the classic discriminant classifier tracking algorithm and the generated model tracking algorithm respectively,and make a reasonable f.usion of the tracking algorithms of the two models.The algorithm in this paper is tested in multiple video sets,and the tracking success rate,accuracy and stability have been greatly improved.Specific innovations and details are as follows:1)For the compression-aware tracking algorithm,the target model and the new target tracking algorithm with a fixed learning rate greatly apply to the stability.In this paper,Bhattacharyya coefficient is introduced to update the learning rate of the target appearance discriminant model.At the same time,when the target is severely obstructed or the attitude is changed frequently in the online tracking of the tracking algorithm,we apply Sigmiod function to construct the model and the new control mechanism,which provides a more reliable and flexible Updated basis.Experimental results show that this method can adapt to the change of the target itself and still be able to effectively track the target when it occludes.2)In order to solve the problem of low tracking accuracy based on L1 minimum tracking algorithm,this paper introduces a more expressive PHOG feature,which contains the spatial characteristics of the image.This reduces the ambiguity of the target expression.At the same time,we segment the target sparsely and determine the occlusion degree of the whole target according to the number of occluded image blocks in the target to determine whether to update the target template.Experiments show that the improved algorithm has higher accuracy of tracking and stability of tracking.3)In order to obtain accurate,efficient and robust tracking performance,this paper combines the tracking algorithm based on discriminant classifier model and the tracking algorithm based on sparse representation generation model,and pioneered the mechanism of embedded fusion,the Integration mechanism effectively integrates the simple and efficient discriminative classifier algorithm and the accurate and stable characteristics of the generated model tracking algorithm.The two algorithms complement each other.By combining the improved compressed sensing tracking algorithm with the sparse representation tracking algorithm,Collaborative target tracking.Experimental results show that the proposed fusion algorithm has better accuracy,stability and tracking efficiency than the existing classical algorithms.