Multiple Human Detection And Tracking In Intelligent Surveillance Video

Intelligent surveillance has attracted significant interests in security monitoring field.The main functions of intelligent surveillance include automatically analyzing surveillance videos,recognizing targets such as pedestrian,human,and vehicle,and storing these relative information for further inquiry.One of the most important functions is multiple human tracking,which has been a classical problem in computer vision and has great potentials in recent artificial intelligence applications such as autonomous driving.In this paper we study the multiple human tracking problem in surveillance videos,proposing a novel pedestrian detection method for better performance in wide-angle surveillance videos,and a detection-based multiple object tracking algorithm which utilizes the structural constraints between tracked objects.In this paper,we employ tracking-by-detection framework for multiple human tracking,which naturally split our works into two parts: detection and tracking.For the detection part,we propose a head detector based on deep learning method,which is the most successful machine learning method recently.As for the tracking part,we propose to utilize structural constraints between tracked objects and detections considering their similarities in multiple dimensions.Another point to be noted is that we discuss the multiple human tracking problem under wide-angle surveillance setting,which brings more difficulties for detection and tracking by less object pixels and ambiguous object features.The novel deep learning based head detector we proposed has been proved to have better performance under the wide-angle setting because it only focuses on human heads,and the tracking algorithm we proposed has a good performance because it relies mostly on structural constraints rather than object features.