A feature fusion pedestrian detection algorithm

When pedestrians are in different angles and positions, the feature extraction and fusion capabilities of YOLO series models are often limited. Aimed at this problem, we propose an improved feature fusion pedestrian detection algorithm YOLO-SCr. To enhance the ability of cross-scale feature extraction and detection speed, we reconstruct the network structure of the YOLO algorithm in the backbone part and convolution layer part, respectively. Then, to strengthen the feature fusion ability of pedestrians at different scales, we introduce the spatial pyramid pooling module and shuffle and Convolutional Block Attention Module (CBAM) attention mechanisms in different positions before YOLO layers. The experimental results show that compared with the detection algorithm such as YOLOv3, YOLO-SCr can effectively improve the detection accuracy, recall and speed.