Acquisition of pedestrian trajectory using on-board monocular camera

This paper presents a general algorithm for pedestrian detection by on-board monocular camera which can be applied to cameras of various view ranges. Under the assumption that motion of background can be nearly approximated as a linear function, the Spatio-Temporal MRF(S-T MRF) model extracts foreground objects as a Region of Interest (ROI). As the foreground objects contain both pedestrians and non-pedestrians, those ROIs are verified by a cascaded classifier. However, the segmentation results sometime do not exactly match with pedestrian on image due to inconstant motion of human leg or merged motion with urban objects. We improved the performance of our system by contaminating errors included in foreground objects using sequential information of ROI. This feedback algorithm drastically decreases the false negative rate. Also, pedestrian trajectory was acquired with high accuracy by referring geometric coordinates. We expect the trajectory could be used as a useful source for measuring the possibility of collision with pedestrian.