부산대학교 도서관

In this paper, we present an ensemble framework with hierarchical and feedback mechanism for object detection. The proposed method is mainly composed of three phases: coarse detection, fine detection and tracking filter. In coarse detection, moving foreground can be rapidly extracted by improved ViBe background subtraction algorithm. FPDW as a fine detector scans the foreground image area, not entire the image, to determine the precise location and the number of targets. In the tracking filter, the detection results are processed to generate trajectories by the Kalman filter. And the current and the next status of the pixel is fed back from the former phases. For assessment of the effectiveness, we implement the proposed framework into pedestrian counting method. Several experiments are carried out based on the benchmark datasets. Experiment results show that the ensemble framework can achieve better detection results and real-time execution in comparison with other the state-of-art methods.