Research On Methods For Pedestrian Tracking Across Non-overlapping Cameras Based On Online Learning

In recent years, pedestrian tracking across non-overlapping cameras is an importantdirection in the research of video moving targets analysis technology. Throughpedestrian tracking in single camera and pedestrian association across multiple cameras,we finally get the trajectory of pedestrian in the entire monitoring network. Thedifferences of pedestrian appearance features under natural conditions is one of the keyand difficult problems. This paper studies a method of pedestrian tracking acrossnon-overlapping cameras based on online learning. Aiming at the difficulties faced bypedestrian tracking across non-overlapping cameras, a basic framework which has thefunction of online learning is proposed. The thought of online le arning is applied toachieve pedestrian tracking in single camera, solving the problem of “drift” very well. Ahierarchical feature pattern is constructed and a association classifier is built to associatepedestrians across multiple cameras based on the samples accumulated in the process ofsingle camera tracking. The main work of this paper includes:(1) A KMD method is proposed to track pedestrian in single camera. Theframework of KMD consists of the Kalman filtering tracker, random fern pedestriandetector and the pedestrian model, making the tracker and detector be complementaryand enhanced to each other to solve the problems of target lose due to occlusion or otherreasons. In the process of tracking, the KMD method learns from samples in the way ofsemi-supervised online learning, making it be able to adapt to the changes of theenvironment and object. The experiments show that the KMD pedestrian trackingmethod achieves a good effect.(2) A hierarchical feature pattern that combines low-level description andhigh-level statistic is proposed to extract pedestrian features, resolving the prominentcontradiction of the invariance and identifiability of pedestrian features. Firstly, we useCSIFT to descript the local feature of pedestrian extracted by intensive sampling.Secondly, Bag of Words approach is adopted to re-organize the low level features, andfinally we obtain CSIFT-BOW pedestrian feature that eliminates the difference ofpedestrian appearance in multiple cameras to some extent.(3) A classifier based association method that utilizes SVM classifier to learn fromthe accumulated samples is proposed. Making full use of different appearance featuresshowed by pedestrian in the process of movement under single camera, this methodprovides an abundant information for pedestrian association in multiple cameras. TheSVM classifier searches for division surface of pedestrian in CSIFT-BOW feature space,and then associate relevant pedestrian in other cameras. The experiments show that thismethod can achieve a high matching degree.