Pedestrian Detection And Tracking In Surveillance

Pedestrian detection and tracking in surveillance is an important research issue of computer vision and pattern recognition, currently. Due to increasing requirement of security in public places, more and more human and financial resources from many well-known companies and research institutions pay for the research of the intelligent monitoring system. This topic can be predicted to have broad prospects and to bring about tremendous social benefits in future.This paper constructed a video sequence based pedestrian detection and tracking system, which included four parts: moving object detection, object tracking in single camera, face detection and pedestrian tracking in camera networks. First, movement information was used to segment foreground regions, which would be tracked in video sequences from one camera. Then in order to determine whether foreground regions were human, face detection was adopted. Finally, integration of human body and face area from multiple cameras could track pedestrian in the network of cameras.In moving object detection, Background subtraction algorithm was researched for both the static and dynamic scenes. Single Gaussian model for indoor environment (static scenes) and Mixture of Gaussian model for outdoor environment (dynamic scenes) effectively completed the task of moving object detection. Color model was adapted to solve the problem of shadow.For single camera, three tracking algorithms were compared: tracking based on connected component, tracking based on Mean Shift and tracking based on Particle filter. A fusion algorithm was proposed: absence of object occlusion, the rapid algorithm based on connected component was adopted; appearing occlusion, particle filter with Mean Shift was used, which greatly reduced the processing time and improved tracking accuracy.In order to determine whether the tracked targets to be a human and position the face region for face recognition, the face detection algorithm was researched. Different from the traditional face detection, due to the small human body in video surveillance, this paper detected face using partial information of upper body, including face, shoulder and head. Harr features from the upper body were extracted and Adaboost algorithm was used for the training and testing.This paper simplified pedestrian tracking in camera networks to the problem of human recognition. As human body image in video monitoring was vague, the performance of popular recognition algorithm decreased greatly. This paper presented a new algorithm: first of all, the face Gabor features and the color histogram of clothes were extracted, and then the similarity of object and the candidate was calculated, human recognition was completed by SVM through the integration of the similarity of the two characteristics.Experimental results show that the framework proposed in this paper can effectively complete the tasks of pedestrian detection and tracking.