Study Of Monocular Pedestrian Tracking Method

Monocular multiple object tracking is essential for computer vision and videoprocessing and becoming more and more important in recent years. If the object isspecialized to pedestrian, monocular multiple object tracking problem will betransformed to monocular pedestrian tracking problem. It can be used in the task ofvideo surveillance, driving assistance, intelligent robot and so on. However, it isextremely challenging. Contrasted to single object tracking, it is more complicate.Moreover, unlike the multiple object tracking based on3D, the occlusion in monocularpedestrian tracking is inevitable.Most of multiple object tracking methods are in tracking-by-detectionframework. With the help of detection, the tracker can automatically deal with the exitand entrance of objects. These methods can be roughly divided into two categories. Oneis based on local strategy, which tracks the objects frame-by-frame. Another is based onglobal strategy, which finds the globally optimal solution for all objects in a largesequence window.Network flow model is a popular graph-based model in the global strategy. Theglobal data association problem is usually mapped to a network model and solved by anoptimization algorithm. To make the tracking result to be more precise, the similarityevaluation between detections is reformulated to adapt the network flow model tocomplex scene. This thesis shows that a greedy optimal algorithm, which has beenproposed to reduce the computational complexity, is not suitable for our network flowmodel. And an improved greedy optimal algorithm is designed to be not only efficientbut also precise.Moreover, a new network flow model based on particle filtering is constructed.This method fuses the local and global strategies to effectively overcome the problemsof tracking-by-detection and it can be divided into two stages: local stage and globalstage. In the local stage, detection responses obtained from the high-quality object detector are combined by particle filter to generate reliable tracklets. In the global stage,the general network flow model is employed to solve the data association betweentracklets. Afterwards, a double-step optimization is proposed to address long termocclusion.