Road obstacle detection is a crucial component of automatic driving environment perception. To enhance the precision of existing road obstacle detection algorithms, we propose an improved YOLOv5s road obstacle detection algorithm. The improved coordinate attention module filters invalid information from multi-scale feature maps and strengthens the focus on areas of interest. Additionally, the enhanced downsampling module alleviates the loss of essential information during sampling in the fusion network, thereby increasing feature robustness. The optimized algorithm′s regression loss and wise gradient gain allocation strategy improve the contribution of common mass anchor frame loss. Experimental results demonstrate that the improved model′s average accuracy on the dataset has increased by 4.2% to 78.6%, outperforming Fast R-CNN, YOLOX, YOLOv7, and other algorithms. With a detection speed of 42 frames per second, the algorithm meets real-time detection requirements. Therefore, the proposed improved algorithm in this study can effectively improve the accuracy of road obstacle detection and has practical application potential.