Moving Target Tracking In Multiple Cameras

Visual monitoring system has been widely used in the field of public security,national defense security and military security.Plenty of challenges as it facing,the visual processing system for target monitoring is still in the low development level of manual retrieval comparison analysis.Because of single camera vision surveillance system deployed in the multi camera perspective and switching illumination variations and differences in appearance,interlaced between dynamic target and frequent occlusions will cause a great impact on the performance of the tracking system.For the reason of that,the target monitoring system based on multiple cameras has great academic value and application prospect.In this situation,this thesis puts forward the solution of multiple cameras and multiple objects tracking,which realizes a high precision of reliable association and continuous tracking.Target tracking in single camera is the basis for multi-camera target correlation and continuous tracking subsequently.Although researchers and related institutions make a great development of single target tracking,the appearance variation due to the difference of illumination and scaling changes lead to tracking drift in the actual environment.For that case,the improved particle filter tracking algorithm based on color difference compensation is proposed in this thesis.The color difference compensation based on the pixel statistical information is transformed into the Lab color space.And the particle tracking is improved by using the gray invariant feature and the MCMC resampling to avoid local optimum.Accurate tracking result also provide reliable observation for object association.Throughout the preprocessing of color compensation,the appearance of the samples captured from different cameras varies due to the pose and shape changes from the perspective transformation,which engender enormous challenges for the object association based on appearance model.In this thesis,we propose the online discriminative appearance model,using MIL algorithm to extract the sequence of samples and multiple feature fusion to descript target.Besides that,we also exploit the MIL-Adaboost algorithm to implement object correlation and identification.Experiments show the accuracy and robustness of the proposed algorithm.Focusing on the ID-switches from the frequent occlusion of overlapping FOV,the object association algorithm based on spatio-temporal and geometric constraints in the overlapping FOV in this thesis.The algorithm takes advantage of spatio-temporal context and the homography relationship between multi-views.With the experiments,the algorithm can estimate the position of occluded targets throughout the occluded region effectively.Experimental evaluations demonstrate excellent performance by reaching the MOTA to 83%,which realize the state-of-the-art level.At the last of thesis,the research work of this thesis is summarized and the further research directions are given.