Abstract
People are the most important objects needed to be detected and monitored in several applications such as surveillance security monitoring, human computer interface, video conference, and virtual reality. Tracking of people becomes a necessary process in surveillance security applications since all significant tracking information of people can be handed over to a higher level of processing, such as human identification, human activity analysis. In this thesis, we address the issues of automatic tracking of humans in video sequences and present our algorithm for tracking multiple people, which has the advantage of tracking recovery from both partial and total occlusions. The recovery process uses a person's model to re-identify and resume tracking of the same person. In order to re-identify the occluded person efficiently, an accurate person's model is needed. One way to obtain the model is when a person's appearance is clearly isolated. Our isolated appearance detection algorithm is also presented. To ensure having an efficient handoff technique among multiple cameras, object identifications across cameras must be consistent. We present our color appearance adjustment for people identification across multiple cameras without color camera calibration or training. The extensive experiments demonstrate and verify efficacy of the proposed algorithms.