Anchor-free pedestrain detection model with semantic context of traffic scenario

Pedestrian detection is an important and challenging issue for autonomous driving. Most of the pedestrian detecting methods utilize the general object detection framework, which follows the two-stage or one-stage pipeline to detect the pedestrian. Nevertheless, these methods usually define the fixed size anchors according to the statistics of the dataset. In this paper, we propose an anchor-free pedestrian detection model. Our model considers pedestrians' semantic context in the traffic scene, which contributes to improving the robustness for small-scale pedestrian detection. Our paper's contributions are: (1) We propose an anchor-free detection network that integrates the segmentation feature. (2) We add an attention module to the network to improve the robustness of detection and make the training process more manageable. (3) We conduct experiments on CityPersons dataset and compared the detection with some state-of-the-art algorithms. Experimental results demonstrate that our algorithm achieves a significant improvement.