Edgelet based human detection and tracking by combined segmentation and soft decision

Human detection and tracking from video stream is important for many applications. Existing detection methods were based on skin-color segmentation or gray level face detection to detect the human. In this paper human detection is based on a silhouette oriented feature called edgelet feature. The system automatically detects and tracks possibly partially occluded humans from a single camera, which is stationary. The discriminative classifiers of objects of a known class are learnt and applied to the video sequence frame by frame. The output of the detection module is a soft decision which consists of a set of detection responses of different confidence levels. The combined detection responses provide the observations used for tracking. The responses of a multiple view detection system are taken as the observation of the human hypotheses. Trajectory initialization and termination rely on the confidences computed from the detection responses. Finally the human is tracked by mean shift style tracker. The system tracks human with interobject and scene occlusions with static or non-static backgrounds. Edgelet features are suitable for human detection as they are relatively invariant to clothing differences.