Automatic Video Object Segmentation and Tracking

Abstract

The segmentation of moving objects with unknown motion under a non-stationary camera is a difficult problem because the camera and object motions are disparate. In addition, the uncertain motion of typical surveillance targets, e.g. slow moving or stopped, abrupt acceleration, and uniform motion makes a single salient motion detection algorithm unsuitable for accurate tracking. This thesis solved this problem by blending the information from the image registration, the frame differences, motion-based segmentation and the spatial segmentation in a non-declarative approach. The image registration is used to generate a motion compensated current frame. Next, temporal differencing and adaptive Kalman filter motion detections are applied to detect the changes in the compensated frames. Finally, detected changes from two motion detection algorithms and the spatial segmentation are combined to identify the moving regions. Experimental results comparing the proposed and other competing methods are evaluated objectively in various accuracy metrics and show that the proposed method achieves promising motion results for a variety of real environments.