The Research Of Prediction Object Tracking Algorithm Based On Compressed Sensing

Object tracking has been an active research topic which has a wide application in many computer vision tasks such as intelligent video surveillance systems, human-computer interaction, virtual realization, intelligent transportation, and so on. Compressed sensing accords with reduce the sampling frequency and reduce the target tracking algorithm complexity. Therefore, it has been received extensive attention in object tracking.The selection of sensing matrix is very important to object tracking algorithm based on compressed sensing. If you select the sensing matrix with highly sparse and can accurate characterization the original features, It can improve the tracking accuracy and reduce the running time. At the same time, the object tracking algorithm based on compressed sensing is tracking-by-detection. It could further improve the tracking accuracy if join the target prediction position in this algorithm framework. But studies of these aspects still are very few so far.It greatly reduced the size of matrix and the running time of the algorithm by using the block compressed sensing matrix which make the scale of sensing matrix independent on the original signal. According to the continuity characteristics of object tracking problems, this paper presents a classification method using prior knowledge to distinguish between object and background through calculation the positive and negative samples prior probability thus increases the difference between object and background. Moreover, in order to avoid the edge information in tracking rectangle interference object tracking algorithm, using the feature extraction method with weighted block which weighted image block with the Gauss distribution and then weakening the background interference.Present the object tracking algorithm by combining the target predicted position with compressive tracking. Put the distance weighted on target predicted position which predict by Mean shift and detect position into classifier. In this way, it could increase the distinguishing between candidate targets and weakened the deceptively to candidate targets and then improve the tracking accuracy. Considering the problem of drift in object tracking algorithm, put forward a kind of nonlinear parameter learning strategies in order to ensure the algorithm still can accurately track the target even in missing or occlusion of the object. It means it will learn the current parameter with larger weight in the case of the target is not lost and learn the prior experience larger if the target is lost. Through this method, it greatly enhance the robustness and anti-jamming of object tracking algorithm because it ensures that the algorithm can quickly learn new knowledge and keep the original model in target loss situation.The results show that the proposed method can overcome occlusion, deformation and other difficult issues in object tracking, and improve the accuracy as well as good performance in real-time tracking by the experiments of multiple video sequences and compared with several state-of-the-art algorithms.