A study of the impact of HOG and LBP based temporal association on far infrared pedestrian detection

In this paper we focus on the problem of pedestrian detection in low visibility conditions, with infrared cameras. Widely applied, tracking is essential for driving assistance applications, providing support for removing false positives and forcing the detection of border line true positives. We propose a multiple feature and temporal based pedestrian detector for far-infrared images. Our model benefits from a fast pedestrian candidate selection given by a three dimensional filtering of bounding boxes that are on the road, it gathers the accuracy of aggregated channel feature pedestrian detector and it is enhanced with a temporal based reasoning mechanism that allows an accurate identification of real pedestrians in the scene. Within the context of the proposed model we study the effects of different association rules for the detected pedestrians. The evaluation of the proposed algorithm proves its benefits, focused on the reduction of the false positives rate.