Research On Tracking Algorithms Based On Collaborative Feature Learning And Regular Correlation Filtering

Target tracking is one of the most important research directions in computer vision,which involves the frontier knowledge of computer science,mathematics,visual cognition and other disciplines and exists in a lot of applications in real life,such as human-computer interaction,video surveillance and so on.After decades of development,there are a large number of excellent target tracking algorithms.However,faced with the complexity of the environment and the diversity of target motion,it is still a very challenging problem to develop robust tracking algorithm.According to the development,target tracking algorithms can be divided into generative method and discriminant method.Generative methods include typical tracking methods,such as particle filter,Mean-shift.With the rise of machine learning,discriminant methods have also received widespread attention,and the typical one is the tracking method based on correlation filtering.On the basis of a careful study of existing target tracking algorithms,two improved algorithms are proposed to overcome some shortcomings of the algorithm.The research contents and achievements of this paper are as follows: 1).Target model is the key of generative tracking algorithm.After analyzing the existing updating strategies,a collaborative feature learning for accurate tracking is proposed.In the framework of improved particle filter,multi-mode collaborative feature learning joints correction mode,updating mode,matching mode and reverting mode to learn the trusted target model,and the appropriate target model is adaptively selected for each frame.Even if the tracking fails,the error can be corrected slowly.Experiments show that the algorithm improves the accuracy of target tracking.2).Since most of the current target tracking algorithms belong to short-term tracking algorithms,a long-term tracking algorithm based on regular correlation filtering is proposed in this paper.Firstly,PCA-like is used to reduce the dimension of fusion features(HOG,CN,integral channel).Then,approximately simplified SRDCF and a proposed supplementary model Color Map are combined to improve the accuracy of short-term tracking.Finally,Proposal detector is proposed to form long-term tracking with short-term tracker.Experiments show that the algorithm has competitive performance in long-term tracking.