Research On Target Tracking Algorithm On TLD Framework

In recent years,with the rapid development of computer vision and artificial intelligence,computer vision has played an increasingly important role in different fields.For example,it has great advantages in identity authentication,missile guidance,military activities and security monitoring.As a very convenient way for human beings to understand outside information,images are closely related to human life.The ultimate goal of the development of computer vision is that computers can understand the image content according to the objective environment and thus make decisions instead of human.Object tracking as an important branch of computer vision,has also attracted people's attention.Until now,there have been many excellent tracking algorithms,but these algorithms often face unexpected difficulties in practical application,such as target deformation,illumination variation,occlusion,scale variation and motion blur.This paper focuses on the study of the occlusion of target and illumination variation in the process of object tracking.It focuses on the exploration of Tracking-Learning-Detection(TLD)tracking algorithm and describes the tracking module,detection module,learning module and synthesis module in the TLD framework in detail,and we analyzed the advantages and disadvantages of TLD tracking algorithm,improved the tracking performance of the original TLD algorithm.This article mainly completed the following aspects of the work:(1)Aiming at the problem of inaccurate tracking when occlusion occurs in the process of tracking of TLD algorithm,the Circulant-Structure-Kernels(CSK)tracking algorithm is introduced in this paper,and a new improved TLD algorithm based on CSK tracking is proposed.The improved TLD algorithm has a template for moving object in the tracking module,it can use the properties of cyclic matrix structure to match the template quickly,and then the point with the largest impulse response of the template can be found in the frequency domain,this point is the target location tracked by the tracking module.The improved algorithm is more accurate than the original TLD tracking algorithm when the target is occluded.(2)After improving the tracking accuracy when the target is occluded,the occluded target sample will enter the sample library of detection module,resulting in a decrease in detection performance of the detection module.To solve this problem,in this paper we combine the perceptual hash algorithm to increase the accuracy of learning samples in the learning module.When the target is occlude,the tracking module traces the object,but it also makes it a positive sample.By using a perceptual hash algorithm,the tracked object can be encoded and the hash code of this object can be obtained,then the hash code is compared with the sample library in the detection module to determine whether the occluded target can be used as a learning sample.Not only improves the performance but also maintains the accuracy of the sample in the sample library.The improved TLD framework has greatly improved the detection performance.(3)Aiming at the problem that the TLD tracking algorithm is vulnerable to the illumination variation in the process of feature extraction.In this paper,the improved HOG feature is introduced,the improved TLD feature extraction operator is no longer obtained by comparing multiple pixel values.Instead,the gradient of each direction of the moving target area is extracted and represented by histogram of oriented gradient,the values of gradient are used to obtain the feature descriptor.The improved TLD feature descriptor has good illumination invariance,and the detection is still accurate when illumination variation occurs.