Research And Implementation Of Pedestrian Tracking System In Intelligent Video Surveillance

In recent years,China has set up the "safe city","smart city" projects while the technologies of digital collection and data storage have benn improved rapidly,as a consequence,image acquisition device has been rapidly popularized,and video surveillance system has become the key component of social public security monitoring platform of city.As one of the key technologies of the monitoring system,the object detection and tracking technology has become one of the hot research topic in China and even the world.In practice,due to some uncertain factors such as the illumination change,morphological scale and strong background interference,occlusion issue,and object loss,the working environment of target tracking system is complex in most situations,which result in higher requirements for the robustness,accuracy and real-time performance of the target tracking algorithm.As a result,the thesis will discuss the above uncertain factors and propose the corresponding solutions to improve the robustness and accuracy of the object tracking system.The main contents of this thesis include the following three aspects:Firstly,we analyse illumination change and part-occlusion issue in the working environment,and propose a generative model tracking algorithm based on sparse representation.Meanwhile,two kinds of particle preprocessing method is proposed,to improve the tracking speed and to reduce the time complexity,while maintaining the tracking algorithm accuracy.Then,we employ local features thought to block the candidate targets,and obtain the local characteristics including part of the overall location information(sparse histogram).Furthennore,we obtain the occlusion state matrix of every candidate target,which can be used to deal with the part occlusion issue,based on the reconstruction error of blocks.Simulation results show that our proposed algorithm is robust to illumination changes and short-term part occlusion.Secondly,we analyse the issue of morphological scale,strong background interference,low tracking error,and low tracking accuracy in working environment,and then propose a generative and discriminative collaborative model tracking algorithm based on sparse representation.In the discriminative model,we try the feature selection of original gray feature to gain the feature of high degree of differentiation.Then,we apply the classifier based on sparse representation to obtain the candidates' confidence coefficient.In the generative model,on the one hand,we employ the kd-tree to gain initial template dictionary set--D rapidly;on the other hand,the sparse structure property is proposed after obtaining the candidates' sparse coefficient vector pool according to the template set--D;then,the sparse coefficient set is obtained according to the sum of blocks' coefficient at the same location of different template in the sparse coefficient vector pool of candidates;at last,the similarity and the total similarity of object candidates is obtained by diagonalizing the sparse coefficient set In the part of template updating,updated template is obtained by applying principal component analysis method,and then,the template which is waiting to be updated in template set is acquired by incremental subspace analysizing,at last,the updating of generative model is finished.After the above processes,the algorithm performs good robustness and tracking performance when dealing with the problems of illumination change,short-term occlusion problem,morphological scale change and strong background interference.At last,an algorithm to deal with object loss problem in the process of object tracking is proposed by analysing the object loss issue in the working environment.The algorithm is consisted of two parts,the one is the process of object tracking,in which the object loss decision step and multi-scale detection part are added to form a complete algorithm.In the object loss decision stage,it can be classified according to the block number and dimension of template dictionary while considering the time complexity of the algorithm.When the partition number is not that much,Bhattacharyya coefficient method is employed to calculate the similarity,and then to judge the object loss by combining the above calculation result and the Noisy-OR model.On the contrary,similarity is calculated acoording to the histogram crossing method,and then,the corresponding threshold to make a decision about object loss can be set.In the multi-scale detection part,the ACF detector proposed by piotr is employed,and then,target in the input image frame can be detected rapidly according to the fast-pyramid feature method and AdaBoost detection model.Finally,the target loss decision and redetection can be achieved.