Obstacle detection affects navigation and obstacle avoidance of autonomous vehicles, where negative obstacles are easy to be missed and falsely detected due to their variable shapes, varying depths, and complex environments. In this paper, we propose a SIP-YOLOv5 network to detect negative obstacles, specifically by adding a small target detection layer and improved coordinate attention to YOLOv5, and incorporating a prediction box correction algorithm to locate and identify negative obstacles. Experimental results show that SIP-YOLOv5 improves AP0.5 by 5.4% and AP05:0.95 by 1.7% compared to YOLOv5m, and it also detects road negative obstacles better compared to methods proposed by others.