Research And Application On Pedestrian Tracking In Video Surveillance

With the development of society,the demand for security is increasing day by day.Public places like suqares are always the focus of security beacuse of its large human traffic.As the limit of traditional surveillance cameras' view,in order to achieve effective monitoring,large numbers of cameras are needed,which not only increase the transmission storage costs,but also increase the workloads of the staffs.Camera system with a fixed camera and a PTZ camera can use a small amount of equipment to balance the breadth and clarity of the monitoring,but the existing camera systems are not satisfied due to its simple tracking algorithm.In order to achieve effective monitoring,this thesis has focused on the research on pedestrian detection and multi-object tracking in videos captured by the fixed camera and singleobject tracking in videos captured by the PTZ camera.In this thesis,we study on the pedestrian detection by aggregated channel features.We determined the final aggregated channel features used in pedestrian detection by many experiments with the Caltech dataset.We make our own training and test dataset to decrease the false negatives rate.In the multi-object tracking process,we use Kalman filter as the filtering algorithm and Hungarian algorithm as the data association algorithm.The assignment cost matrix of the Hungarian algorithm is not only computed by the position information but also the appearance information.As the multiple object tracking accuracy cannot satisfy our requirements,we then study on single object tracking algorithm.TLD tracking algorithm is chosen and improved to track nonrigid object like pedestrian.At last,we make a tracking experiments in the dynamic scene and come to the conclusion that TLD tracking algorithm can meet our requiements.