ESD-YOLO: a lightweight adaptive multi-scale object detection algorithm for autonomous driving

Abstract Real-time object detection during driving plays a vital role in the driving decision-making of unmanned vehicles. To meet the demands of lightweight design and real-time performance, this paper proposes an improved multi-object detection algorithm for unmanned driving scenarios, ESD-you only look once (YOLO), which is based on YOLOv11. In ESD-YOLO, the C2PSA module in the backbone network is replaced by the multi-scale enhancement block (MSEB). By virtue of its multi-branch structure and dynamic parameter generation, MSEB ensures both lightweight performance and detection accuracy. The head structure is modified, with the original detect module replaced by enhanced light detect (EL-Detect). EL-Detect achieves a balance between the lightweight nature and accuracy of the detection head through a lightweight stem structure, the embedded simple and efficient attention module attention mechanism, and a lightweight prediction head structure. Moreover, the original loss function is substituted with the improved adaptive SIoU (ASIoU). Based on the ordinary SIoU loss function, ASIoU enhances the convergence of small objects by adding target scale adaptive weights and reducing the shape cost weight for small objects, thereby compensating for missed detections in multi-object detection. Experimental results demonstrate that the algorithm achieves a mAP@0.5 of 96.1 % on the Kitti dataset, which is 2.2 % higher than that of the original YOLOv11s network. Meanwhile, the number of parameters is reduced by 31.8 % , significantly improving the model’s ability to detect multiple objects in unmanned driving scenarios, and its performance is superior to that of other algorithms. The generalization of the model is verified through experiments on the BDD100K dataset.