Research On The Robustness Of Compressive Tracking Algorithm For Video Object

With the popularization of computer applications and the development of image processing technology,video target tracking has succeed in attracting more and more attention,which has great significance in traffic management,intelligent monitoring,human-computer interaction and smart cities.Under the multi-disciplinary impact of image processing,cybernetics and control engineering,computer science and pattern recognition,there are many excellent target tracking algorithms arised in the field of computer vision.Recently,compressive tracking algorithm has become a research hotspot due to outstanding performance and faster calculation speed.Although compressive tracking has its own unique advantages,when it comes to such factors as occlusion,multi-scale variations,abrupt motion,the similarity between the foreground and the background,it is easier to drift,which resulting in tracking failure.Therefore,robustness optimization for compressive tracking has important theoretical and practical value.This paper has fully investigated the results of previous studies,and put forward corresponding improvement measures for major issues based on the current-status of research.The main research work is as follows:1.We propose an improvement for classical compressive tracking based on fragment weighting.The algorithm can effectively solve the problem of tracking task failure when the moving target is occluded.The algorithm adopts the idea of block weighting,which divides each frame into blocks,then makes a similarity judgment and establishes an occlusion judgment mechanism to ensure the necessary real-time performance.Once the target is occlued,the normal sub-block plays an important role in determining the candidate target position of the current frame,which plays a central role in this algorithm.Then the positive and negative samples are taken near the candidate target position.After the sparse matrix compression,the Na?ve Bayes classifier is used to obtain the predicted position of the next frame.Experiments show that the proposed improved method can track the target more accurately,and has better reliability and stability than the classical algorithm.2.2.We adopt a course-fine strategy and put forward an improved algorithm based on spatiotemporal context and particle filter.The algorithm can well deal with multi-scale variations when the target is under the fixed motion scenes.The algorithm firstly uses particle filtering method to obtain the rough position of the current frame target on the previous frame in time series.In the spatial context of the current frame,we use the method of resizing the actual tracking box,which means changing the scale of the actual tracking frame,and then compare with the previous coarse position respectively,we take the corresponding tracking frame with the highest similarity as the positive sample of the final target candidate position.Therefore,take the original features for compression,and finally get the target prediction position of the next frame through the classifier.Such improvement can not only enhance the robustness,but also guarantee real-time performance in application.Through the improvements of the original algorithm mentioned above,the algorithm can effectively continue to track the target when the target has occlusion and scale changes.However,the computational complexity increases while the algorithm improves robustness.In addition,the algorithm is also easy to track and fail when the target is in abrupt motion,or the foreground and background are in clustering.Thus how to ensure that the real-time performance become a research focus of the improvement of compression tracking algorithm in the future.