Temporally integrated pedestrian detection from non-stationary video

In this paper, we propose a novel approach for detecting pedestrians from video sequence acquired with non-static camera. The proposed algorithm consists of three major components, including global motion estimation with motion-compensated frame subtraction, AdaBoost pedestrian detection, and temporal integration. The global motion estimation with frame subtraction can reduce the influence of the background pixels and improve the detection accuracy and efficiency. 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 frame subtraction, the AdaBoost classifier is employed to detection pedestrians in a single frame. At last, the graph structure is applied to model the relationship of different detection windows in the temporal domain. Similar detected windows are grouped as the same clusters by using the optimal linking algorithm. The missed detection windows will be recovered from the object clustering results. Finally, we show the experimental results by using the proposed pedestrian detection algorithm on some real video sequences to demonstrate its high detection accuracy and low false alarm rate.