Research on Vehicle and Pedestrian Object Detection Algorithm based on Enhanced YOLOv5s

Autonomous driving technology relies heavily on the accurate detection of vehicles and pedestrians. The swift and reliable identification of these objects in complex road environments has extensive implications for traffic congestion alleviation and accident prevention. However, current object detection algorithms used in autonomous driving and related fields face challenges such as misidentifying occluded targets, high computational complexity, and low detection accuracy. This paper proposes an enhanced vehicle and pedestrian object detection method based on YOLOv5s. Our approach addresses the detection requirements for small and occluded targets by incorporating the Mosaic-9 data augmentation strategy during the training process. To meet the low computational demands in vehicular environments, we replace the convolutional layers in the YOLOv5s algorithm with GhostNet modules. Furthermore, we introduce the Coordinated Attention module after the original YOLOv5s algorithm's feature extraction network to enhance the network's learning capability. Experimental results indicate that our improved YOLOv5s object detection network, compared to the original network, reduces the floating-point computational complexity by 2.1 GFLOPs while achieving an approximate 4% increase in average detection accuracy for vehicles and pedestrians. These findings demonstrate the effectiveness of our approach in enhancing object detection performance in autonomous driving scenarios. (Abstract)