Lightweight Road Object Detection Algorithm Based on LEP-YOLO

Driven by the rapid advancement of computer vision, road object detection has demonstrated significant application value in the field of intelligent driving. To address the challenges of high miss rates and limited real-time performance caused by multi-scale objects and complex background interference in real-world road scenes, this study proposes a lightweight model named LEP-YOLO. The backbone network incorporates Large Selective Kernel Attention (LSKA), effectively enhancing the model's robustness against interference. An Efficient Detail-Enhancement Detection Head (EDE-Head) is designed to improve the detection head's capability for capturing fine details. Channel pruning is employed to reduce redundant model channels, effectively compressing the model size and accelerating detection speed. Experimental results on the KITTI datasets demonstrate that the proposed model achieves 2% improvements in mAP@50 compared to baseline models, respectively. The parameter count and computational complexity are reduced by 54.6% and 32.7%, respectively. The detection speed reaches 228 frames per second (FPS), achieving a balance between detection accuracy and model lightweighting.