Vehicle–Pedestrian Detection Method Based on Improved YOLOv8

The YOLO series of target detection networks are widely used in transportation targets due to the advantages of high detection accuracy and good real-time performance. However, it also has some limitations, such as poor detection in scenes with large-scale variations, a large number of computational resources being consumed, and occupation of more storage space. To address these issues, this study uses the YOLOv8n model as the benchmark and makes the following four improvements: (1) embedding the BiFormer attention mechanism in the Neck layer to capture the associations and dependencies between the features more efficiently; (2) adding a 160 × 160 small-scale target detection header in the Head layer of the network to enhance the pedestrian and motorcycle detection capability; (3) adopting a weighted bidirectional feature pyramid structure to enhance the feature fusion capability of the network; and (4) making WIoUv3 as a loss function to enhance the focus on common quality anchor frames. Based on the improvement strategies, the evaluation metrics of the model have improved significantly. Compared to the original YOLOv8n, the mAP reaches 95.9%, representing an increase of 4.7 percentage points, and the mAP50:95 reaches 74.5%, reflecting an improvement of 6.2 percentage points.