Pedestrian Target Detection and Behavior Recognition of Driverless Vehicles Based on Deep Learning

Traditional methods are difficult to meet the needs of driverless system for pedestrian intention understanding due to insufficient adaptability to dynamic scenes, limited real-time performance and poor accuracy of behavior prediction. In this paper, we propose an improved YOLOv8 multitask deep learning framework (BiFaster-YOLOv8), which integrates FasterNet lightweight backbone network, BiFormer two-layer dynamic routing attention mechanism and Focal-EIoU loss function to achieve collaborative optimization of detection accuracy and inference speed. FasterNet reduces 85% of trunk parameters by local calculation and global transmission strategy, BiFormer attention mechanism combines global coarse screening and local refinement to enhance key feature extraction ability, Focal-EIoU effectively alleviates sample uneven distribution problem. Experiments on the self-built driverless pedestrian dataset show that the mAP@ 0.5 and recall of the improved model reach 89.2% and 86.3% respectively, the inference speed is increased to 63 FPS, and the parameter quantity is compressed to 9.8 M, which is significantly improved compared with the original YOLOv8.