Research On Target Tracking Algorithm Based On Compressive Sensing

Target tracking has become one of computer vision research hot topic. Target tracking are concernedboth in the fields of scientific research and engineering application, and target tracking technology has been widely used in various fields,such as: missile system, aircraft navigation system. In this paper, utilize the idea of the compressed sensing theory and target tracking technology in-depth study, try to find a stable and high efficiency tracking algorithm. The specific research content are as follows:(1) Online multi-instance learning Target Tracking Algorithm based on compressive sensing. First, reduce the dimension of extracted multi-scale images, then use the online multi instance classifier to classify the extracted features, and real-time update the classifier. The combination of sparse representation and multi-instance learning classifier improves the accuracy of target tracking algorithm, and solves the problem of template drift.(2) Real-time Compressed Sensing Tracking Algorithm, the algorithm thought is as follows: Firstly, reduce the dimension of multi-scale image features by using the random measurement matrix which conforms to the compressive sensing RIP condition, obtain a low dimensional compression subspace, then classify the feature after dimension reduction by simple Naive Bayes classifier. The tracking algorithm is simple, but the results of the experiment show that the robustness is well reflected,and its speed can reach nearly 40 frames per second.(3) In view of the target online model updating error causes tracking drift problems in the existing online learning tracking algorithm, this paper puts forward an online adaptive model update target tracking algorithm: firstly, compressed sensing technology is highly efficient, to reduce the dimensionality of multi-scale image features, and extract multi-scale sample to achieve target scale adaptive update, then use extracted positive and negative samples low dimensional image features to train a Naive Bayes classifier, use the classifier output target sample target of maxconfidence to achieve target tracking, and according to the current target confidence to adaptive online model update rate, reduce the target error update caused by occlusion. Experiments show that the proposed method can achieve better robustness tracking in the case of scale change, local and global occlusion, illumination change.The average tracking success rate is improved by 20.3% compared with the original compressed sensing algorithm.