Traditional tracking is composed of trajectory initialization, object following and trajectory termination. In the previous approaches, trajectory initialization and termination fully depend on object detector?s responses and object following is performed by classic tracking approaches, such as Kalman filter, particle filtering and Mean-Shift tracking.
However, in practical environments human detector has relatively lower detection rate and higher false positive error than those of other object detector due to large variations of appearance and pose and complex background such as street scene. Furthermore, classic tracking algorithms are only valid on short time interval due to frequent self-occlusion by highly articulated human body and serious occlusion by other people on crowded street. We are developing a method which can automatically initialize and terminate paths of people and track multiple and changeable number of people on cluttered scenes over long time intervals by combining people detection and local feature tracking.
Our probabilistic graphical model makes it possible to find accurate paths by removing mis-detected patches and recovering temporarily lost patches in global optimization framework.