Pedestrian Detection Based on YOLO-D Network

In view of the poor robustness of pedestrian detector based on artificial extraction feature, a novel pedestrian detection method is proposed by referring to the research results of state-of-the-art object detection. Based on the YOLOv2 network and inspired by the DenseNet, we pass the low-level feature maps of YOLOv2 to the higher layers in turn. By combining the feature maps of different convolutional layers, we propose a novel network architecture, namely YOLO-D network, to make detector performance better. In order to solve the occlusion problem, we introduced a Head-Shoulders model and also reduced computer calculations. The experimental results show that compared with the original YOLOv2 network pedestrian detection method, this paper reduces the missed rate and false rate, improves the localization precision, and the detection speed meets the real-time requirements.