Researches On Object Tracking And Event Detection Based On Tracklet Association

With the development of digital surveillance equipments technology and the cost of these equipments become lower, more and more surveillance system have been implemented in traffic, security area and so on. The continuous working of these systems generate huge amount of video sequence data. How to analyze these data automaticlly is a hotspot which captures much attention. An intelligent surveillance system requires several different video processing technologies in different level, from low level tracking to high level event detection. To address these problems, we invetigated these video processing technologies including object detection, object tracking, surveillance scene modeling and video event detection based on tracklet association technieque.A hierarchical tracklet association algorithm is proposed. Since current object detection algorithms still cannot detect objects accurately and continuously, which results in objects'movement in the scene discontinuously and hardly to track them. Hierarchical tracklet association algortithm can asscotiate those tracklets that belong to the same object to achieve multiple objects tracking. The hierarchical association can reduce the association solution searching space and improve association algorithm's effectiveness.In moving object tracking, we present a multiple objects tracking algorithm base on tracklet association, which can track objects under complex scene where both occlussions between objects and occlusions between object and the scene happen. Through a primitive association between detection response and a higher level hierarchical tracklet assocation, object's tracklet caused by occlusion or misdetection is prograssingly assocaited and object's full trajectory is achieve. Our algorithm can track object when occlusion and detection failure happens, and the hierarchical association by prograssingly associate tracklets with wider gaps can overcome the drawbacks of inefficent searching that other association algorithm have.In moving object detection, we present an enhanced object detection algorithm based on tracklet association. By investage the relationship between object detection and object tracking, we set object detection and object tracking into one optimal framework which combines single frame information by detection and temporal information by tracking. By employ the two level association from tracklet association technieque; we can track multiple objects and improve detection results. And can verify misdetection in the past frames.In surveillance scene modeling, we present a surveillance scene modeling algorithm based on tracklet association. By analying the reason of why an object's trajectory falls into several tracklets and compute these trajectory gaps, we can model a surveillance scene with entry area, exit area, main pathes and scene occluders. Unlike scene modeling technieques based on statistical trajectory result, our method do not need huge amount of trajectories as prior knowledge. we can model the scene with the process of tracking. The knowledge of scene models like entry area, exit area, main pathes and scene occluders can be helpful in improving tracking results and higer level video event detections.In video event detection, we present a primitive event detection algorithm based on trackleta assocation and a abnormal activity detection algorithm based on scene model. We compute the likelihood of primitive event like object enter and/or leave the scene, two objects merge and/or split by Bayesian theory with the knowledge of spatio-temporal information of tracklet gaps. By employ complex event model like Hidden Markov Model to model those primitive events, we can detecte more complex events. By analy the consistence of the object's trajectory with the scene model, we can decide an object's activity is noraml or abnormal. And by computing the likelihood of a object's trajectory oberse sequence with the Hidden Markov Model of the scene, we can have the probobalistic based abnormal activity representation.