Research On Object Tracking In Intelligent Video Surveillance System

In recent years,Intelligent Visual Surveillance(IVS)system has become one of the great hot research area in computer vision.Intelligent Visual Surveillance system used computer vision technology and pattern recognition technology to realize autonomously image processing,image analysis and image understanding in surveillance scene.The core technologies of Intelligent Visual Surveillance system involve: object detection,object tracking,object recognition and behavior understanding.The Intelligent Visual Surveillance system has widely applied in public security management,intelligent traffic management,construction of smart city and many other aspects,which has shown a great effect on social benefit and economic benefit.As one of the key technology in Intelligent Visual Surveillance system,object tracking plays an important role.On basis of finishing the object detection,object tracking technology locates target accurately and efficiently,which provides support to the following work,such as event detection,behavior recognition and so on.For the problems of object tracking in Intelligent Visual Surveillance system,the main work and results in this dissertation are as follows:(1)As the color histogram is easily disturbed by background region or other objects with similar color distribution,a new method based on spatiogram and particle filter tracking algorithm is proposed.Spatiogram is applied to represent the appearance model of object and then object-background similarity weighted Jenson-Shannon Divergence(JSD)measurement is employed to measure the similarity between object region and the candidate region,which strengthens the discriminability of appearance model and improves the robustness of the tracking process.(2)For the problem of occlusion in visual surveillance scene,a novel object tracking algorithm based on Bayesian Decision and Particle Filter is proposed.Based on analyzing the process of occlusion,spatial uncertainty is applied to measure the stability of object tracking and Bayesian Decision method is adopted to judge the object encounters occlusion or not,which decides the template updating strategy adaptively.The above method has been involved in particle filter framework successfully,contributing to high ability of handling occlusion.(3)For the requirement of real-time and accuracy for object tracking in Intelligent Visual Surveillance system,a tracking algorithm named multiple feature weighted Spatio-temporal Context Learning(MFWSTC)is proposed.For one hand,the employment of STC tracker promotes the tracking algorithm's computational efficiency.For the other hand,the usage of multiple features model and the weighted matrix of measuring the contribution of different features enhance the discrimination ability of appearance model.Finally,the Kullback-Leibler(KL)divergence is applied to fuse the confidence map of different features,which ensures the effective information of different features has been guaranteed to be fully used.The large number of experimental results illustrate that the proposed algorithm achieves the promised performance.(4)In the view of the problem of illumination variation,scale variation and occlusion during object tracking in visual surveillance scene,a multi-scale context-aware correlation filter tracking algorithm based on channel reliability is proposed.Firstly,histogram of oriented gradient(Ho G)feature,gray feature and color names(CN)feature are extracted as the appearance model of the object,which can enhance the robustness of the tracking algorithm in complex scene.Secondly,in order to avoid the distribution among all channels,the multi-channel context-aware correlation tracker is independently trained by the related channel feature sample.Then,the channel reliability is established to value the confidence of each channel,which is used to fuse the final response map of the multi-channel context-aware correlation tracker.Finally,the scale pool method is applied to estimate the position and scale of object optimally.The experimental results show that,the proposed algorithm can tackle the illumination variation,occlusion,scale variation and other complicated factors efficiently,which owns high accuracy and robustness.For the problem of visual tracking in Intelligent Visual Surveillance system,several object tracking algorithms based on different complex scene variations are proposed,which laid a solid foundation for the implementation of accurate,robust and efficient object tracking in Intelligent Visual Surveillance system.