BFOP3D: binocular feature based occluded pedestrian 3D detection

3D pedestrian detection is an important problem in most intelligent transportation fields. Pedestrian detection in 3D under crowding scenes is big challenge of most detection models based on deep learning networks. Methods usually used to produce the 3D location of pedestrians could be split into two parts, i.e., 2D pedestrian detection and depth map prediction. The poor coupling between these two tasks results in inferior performance of such methods. In this paper, we propose a light-weight 3D pedestrian detection model based on binocular images which directly generates 3D bounding boxes of pedestrians. It is designed to alleviate the negative impact on occlusion problem by applying multi-scale stereo features fusion and data augmentation strategy. The model we proposed outperforms most 3D pedestrian detection methods on KITTI dataset.