An Enhanced YOLOv8 Algorithm with C2f_Parnet for Vehicle and Pedestrian Detection

This paper proposes enhancements to the YOLOv8s model for autonomous driving object detection tasks. Firstly, a C2f Parnet module is designed based on the Parnet parallel network architecture to reduce network depth and computational latency. Secondly, GSConv is modified by employing dual 3 × 3 convolutions to decrease parameter count. Thirdly, L2-scored structural pruning is introduced to compress redundant channels. Finally, detection accuracy is improved using EIoU loss. Experiments on the KITTI dataset demonstrate that the improved model maintains a mAP50 of 0.917. The performance of this model is almost comparable to the original model (0.939), but with parameters reduced to 11 million and GFLOPs decreased to 26.2. Compared with other methods, L2 pruning achieved the highest mAP50-95 score of 0.753. In complex traffic environments, this model demonstrates robust real-time detection capabilities, can be flexibly deployed, and is a useful choice for lightweight detection in autonomous driving.