Pedestrian Detection Method Based on FCOS-DEFPN Model

Abstract

Automatic driving technology has high accuracy and real-time requirements for pedestrian identification and localization. Pedestrian detection is a basic and necessary function in vision-based pedestrian detection systems and collision warning, which can effectively avoid traffic accidents and improve road driving safety to a certain extent. In this paper, a lightweight solution based on the FCOS-DEFPN model is proposed for real-time pedestrian detection. Based on the FCOS model, this paper proposes the FCOS-DEFPN model, which achieves the lightweight of the network by replacing the ResNet50 backbone network with the MobilenetV3 network and using the depth separable convolution instead of the ordinary convolution for parameter compression. While maintaining the detection accuracy, this paper introduces data enhancement methods such as Random Erasing and Morsia to simulate pedestrian occlusion and small target scenarios to improve the robustness of the model. For the pedestrian occlusion scenario, this paper introduces a lightweight attention network ECA, which helps to extract pedestrian features better. For small-target multi-scale pedestrians, the DEFPN feature pyramid network is proposed, which acquires feature information at multiple scales by attentional fusion of feature layers at different scales from top-down, bottom-up, and front-back. The experimental results show that the proposed model is enhanced in terms of detection accuracy for occluded and small-target pedestrians, and satisfies real-time pedestrian detection under the premise of robustness in complex scenes.

Keywords

• Automatic driving
• pedestrian detection
• full convolutional one-stage target detection
• small target detection
• occlusion detection