Online Multi-target Visual Tracking

Visual object tracking is a challenging task in computer vision. It has manyapplications ranging from video surveillance, human-computer interaction, videoindexing, vehicle navigation to trafc monitoring. Because of the complexity ofbackground, appearance variation, illumination changes and occlusions, visualtracking has many problems to solve. This thesis mainly studies two types ofmulti-target tracking algorithms. One type of methods is those tracking based ondiferent categories, which is called Category Free Tracking. The objects in thesemethods are not known in advance and the only information about them comesfrom the frst frame where they are selected by the user. Then, an online learnedclassifer is utilized to track diferent targets and thus this type of algorithmsare also called online visual tracking. The other type of algorithms are thosefocusing on specifc kinds of objects, which is named Association Based Tracking.They use a pre-trained detector for certain kind of objects to generate detectioncandidates, then associate them into trackers.For the frst type of methods, we propose a novel tracking framework thatadapts various appearance changes of object and also owns the ability of reacqui-sition after drift. A Condensation-based method with an online support vectormachine as observation model is adopted. In order to redetect the object whenthe target drifts, we utilized a detector inspired by random ferns. We also presenta refnement strategy to improve the tracker s performance by discarding thesupport vector corresponding to possible wrong updates by a matching templateafter reinitialization. The experiments shows that our method outperforms othertwo algorithms in BoBoT dataset and demonstrate a promising performance forthe rest of datasets. For the second type of algorithm, We assign a Condensation-based tracker with online learned classifer to each target in order to associate thecorresponding detection responses and thus realize multi-target tracking. Basedon the results of data association, we integrates the target s velocity into weightscalculation to handle object occlusion assuming that fast-moving target is not likely to change directions abruptly because of inertia. In addition, we design anew data association method whose afnity measure is computed by the classi-fer score judged on candidate image patch, the distance and size similarity oftwo rectangles. The experiments reveal that our algorithm has an advantage inMOTP metrics comparing with other methods.