Abstract
In this thesis, we propose an integrated approach that can detect pedestrian system from video sequence acquired with non-static camera environment. The proposed system contains three major components, including global motion estimation with background subtraction, AdaBoost pedestrian detection, and temporal integration. The global motion estimation with background subtraction can reduce the influence of the background pixels and improve the detection accuracy. The simplified affine model is used to fit the global motion model from some reliable blocks by using the RANSAC robust estimation algorithm. After motion-compensated background subtraction, the AdaBoost learning algorithm is employed to detection the pedestrian in a single frame. At last, the graph structure is applied to model the relationship of different detection windows in the temporal domain. The similar detection windows will be grouped as the same clusters by using the optimal linking algorithm. The missed detection windows will be recovered in the clusters containing larger nodes. In the experimental results, we capture three kinds of video in the campus of National Tsing Hua University to evaluate the performance of our system. The experimental results demonstrate that the proposed system achieves high detection accuracy and low false alarm rate.