Year: 2024
Type: conference-paper
Abstract: To address common challenges in small object pedestrian detection, such as difficulties in feature extraction, background noise interference, and underutilization of shallow feature information, this paper proposes an improved road pedestrian detection method based on the YOLOv8 network. First, a novel architecture called Multi-layer Fusion of Shallow Feature Information FPN (MFSF-FPN) is introduced into the neck structure of the YOLOv8 model. This architecture enhances the expression capability of semantic information in shallow networks by integrating shallow networks with partially interrupted bottom-up feature pyramids through multi-scale fusion. Secondly, introducing Polarized Triplet Attention Convolution (PTAC), which optimizes the Polarized Self-Attention mechanism to address the shortcomings of Triplet Attention Convolution in utilizing spatial and channel features, thereby enhancing the ability to extract small object pedestrian features. Finally, using the Feature Adaptively Spatial Feature Fusion (FASFF) detection head by adding extra detection layers to capture feature information at different scales and optimizing the feature fusion process to avoid loss of semantic information during cross-scale fusion. Experimental results demonstrate that this method significantly improves detection accuracy while achieving lightweight performance, with [email protected] increasing from 63.5% to 72.5% and mAP50-95 increasing from 39.7% to 47.8%.
Authors: Yuchen Pan, Wenhua Han, Zhuangzhuang Bao
Institution: Shanghai University of Electric Power
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4409095352