Abstract
Recently, the study of human object tracking has been apparently moved from camera-based representation to object-based representation because object-based representation can assist people to trace the human behavior and search for the abnormal conditions effectively.The thesis proposes an automatic object-based tracking system using distributed multiple cameras system with non-overlapping viewing range. The goal of tracking between multiple cameras with disjoint view is to establish a set of correspondence between observations of objects across multiple cameras. Two visual cues, spatiotemporal cue and appearance cue, are used for tracking human objects across cameras. To learn the relationships among cameras, we use batch-learning procedure and update all probability matrixes constantly for long-term monitoring. Also, we improve the correspondence of appearance cue by color calibration among different cameras. Under certain conditions, the tracking of human objects may be lost due to light variation, unusual behavior, or slightly color change of clothes in different camera views. The proposed work use the dynamic programming algorithm to backward tracking with spatiotemporal relationships to search more information. We can use this information to link the missing linkage in the tracking path. Hidden Markov Models are further used to verify the abnormal path of human object across multiple cameras according the training data. Experimental results show the efficiency of the proposed method.