Pyramid binary pattern features for real-time pedestrian detection from infrared videos

This paper presents a robust real-time pedestrian detection approach from infrared (IR) videos using binary pattern features. A novel pyramid binary pattern (PBP) feature is first proposed for IR pedestrian appearance representation. Both symmetry and spatial layout of texture cells have been encapsulated in the PBP feature. PBP outperforms several state-of-the-art binary pattern features for IR pedestrian images classification. Motivated by the recent success of motion-enhanced pedestrian detector, we then extend the PBP feature to 3D spatial–temporal volumes. The dynamic PBP feature combines both motion and appearance for IR pedestrian description and achieves better performance in comparison to the static PBP feature. Finally, a keypoint based sliding window support vector machine (SVM) classifier is used to detect pedestrians in IR videos. The keypoint based scanning strategy reduces the number of candidate sub-windows dramatically. The proposed approach has been implemented on an experimental vehicle equipped with a forward-looking infrared (FLIR) camera. Experimental results in various urban scenarios demonstrate the effectiveness and robustness of our approach. In addition, even though our approach is presented for IR imageries, it can also be applied to pedestrian detection in visual images.