Research On Visual Tracking And Application Based On Sparse Representation And Compressive Sensing

With the development of electronic,communication and computer technology,video surveillance system will be playing more and more important role in the field of traffic monitoring,community security and inland river shipping.Visual tracking technology is one of the most important part of intelligence video surveillance system,and is a necessary precondition of activity recognition and event handling.In recent years,the sparse representation and compressive sensing was used for object tracking,then several tracking algorithms with high precision or good real-time have been proposed.However,practical experience has shown that real-time tracking technology in complex scene(such as scale variation,illumination change,occlusion,pose variation,rotation,abrupt motion and background cluttering,etc)is far from mature.Therefore,it is still faced with many challenges.In addition,there were few reports about tracking algorithms of inland ship.Hence,research on the inland ship target tracking and handover algorithms with high speed and robustnessis important meaning for the riversurveillancevideosystem.Based on the theory of sparse representation and compressive sensing,the object tracking and handover technology under complex scenes has been researched and explored in this thesis.The main work and achievements are shown as following:(1)Although the compressive tracking(CT)algorithm can work effectively,it may drift away or fail especially when the target is occluded heavily and continuously.To overcome this shortcoming,a real-time tracking method based on sub-region classifiers and compressive sensing is put forward.Firstly,the target region is divided into four sub-regions in a fixed mode,and four sub-region classifiers will be generated subsequently.On the assumption of rigidity,the final location of the target can be evaluated by the sub-region with the highest confidence.Finally,a simple but feasible update strategy is used for these sub-region classifiers.Experimental results show that the proposed method is more robust to occlusion than CT algorithm.(2)L2-RLS based object tracking(L2)algorithm can deal with some complex appearance changes in the video scene successfully,but it is prone to drifting when the target object undergoes pose variation or rotation.To deal with this problem,a robust object tracking method is achieved by combining L2-RLS and compressed Haar-like features matching.Firstly,the extent of occlusion can be evaluated by L2tracker.Secondly,the compressed features matching method is used to locate the target object if the extent of occlusion satisfies two inequality constraints.Finally,most of the insignificant samples are removed by the minimum error bounded criterion,which significantly reduces the computational complexity.Both qualitative and quantitative evaluations on numerous challenging image sequences demonstrate that the proposed method is more robust and stable than L2 tracker when the target object undergoes pose variation or rotation,but its tracking speed is only slightly slower than L2 tracker.(3)Owing to removing the trivial templates in sparse representation,the object tracking algorithm based on L2-norm minimization has good real-time feature.However,it achieves poor tracking performance when the target object undergoes pose variation or rotation.To deal with this problem,a fast visual tracking method is proposed based on L2-norm minimization and compressed Haar-like features matching(FL2CHFM).The proposed method not only removes square templates,but also presents a simple but effective observation likelihood which reduces the computational complexity to a certain extent,and its robustness to pose variation and rotation is strengthened by Haar-like features matching.Compared with other popular methods,FL2CHFM has stronger robustness under complex scenes and runs fast with a speed of about 29 frames/s.Finally,the fusion algorithm under low-resolution(FL2CHFM_L)is applied to inland ship tracking.Experimental results show that FL2CHFM_L shows better stability and higher accuracy in comparison with several fast tracking algorithms.Furthermore,it meets the requirement of real-time ship tracking.(4)Object handover between adjacent cameras is key to keep track of ship monitored by a multi-camera network with non-overlapping views.A new method for fast re-identifying and locating inland ship is achieved,in which the appearance of wheelhouse,but not the whole ship is employed.The main contributions include:(1)The improved fast compressive tracking algorithm is used for detecting and locating the object quickly.(2)The maximum classifier scores is used to locate the time when a ship comes into view.(3)The multi-scale Histogram of Oriented Gradients features are employed to distinguish ships with very similar appearance.(4)The proposed method runs fast with a speed of about 60 frames/s,and meets the requirement of fast ship handover.