GNN-Enhanced YOLOv4 for Pedestrian Detection

The YOLOv4 pedestrian detection method originally efficiently integrated multi-level feature maps through PAN networks, greatly improving the efficiency of information aggregation and transmission. However, the downsampling and upsampling operations in PAN networks during processing may cause information loss and introduce unnecessary noise during upsampling. To overcome this bottleneck, this paper proposes an innovative YOLOv4 pedestrian detection method, which introduces Graph Neural Network (GNN) instead of PAN network. Through GNN, this method can automatically optimize the information aggregation and transmission process, while effectively avoiding the potential problems of downsampling and upsampling, significantly improving the accuracy and efficiency of pedestrian detection. The experimental results show that our method achieved 89.3% AP value and 92.0% AR value on the VOC2007 dataset.