Research on a Road Object Detection Algorithm Based on an Improved YOLOv11

To improve the accuracy of existing road object detection methods, we propose an improved algorithm based on YOLOv11. We replace the C3K2 module in the backbone of YOLOv11 with a CPMSFA module. In the neck part of the network, we use the LeakyReLU activation function and redesign it as an ASFP2 structure. This change helps the network handle features of different sizes better. We also design a new detection head called DDCNV, which uses dynamic convolution and depthwise separable convolution instead of the traditional one. This helps the model detect objects with different shapes more effectively. We test the model on the public KITTI dataset. The results show that, compared with the original YOLOv11 model, our method improves the mAP@0.5 by 4.4% and the precision by 4.5%. These results show that the new model works better than the baseline.