Abstract:A lightweight YOLOv5s-based road-assistance detector, EBD-YOLO, is proposed for resource-constrained embedded terminals. The efficient vision transformer (EfficientViT) is adopted as the backbone to reduce model complexity. Focal-generalized intersection over union (Focal-GIoU) improves bounding-box regression. The bidirectional feature pyramid network (BiFPN) enhances multiscale feature fusion. Dynamic head integrates attention into the detection head, and Soft-complete intersection over union non-maximum suppression (Soft-CIoU_NMS) improves occluded-object detection. Experiments on KITTI and BDD100K datasets show that the model size is reduced by 38.4% and 37.2%, respectively, and computational cost by 48.1%. Meanwhile, the mean average precision at IoU threshold of 0.5 (mAP@0.5) increases by 0.5% and 5.8%, and the mean average precision at IoU threshold of 0.5:0.95 (mAP@0.5:0.95) by 2.8% and 7%, respectively, demonstrating the feasibility of deployment on resource-constrained in-vehicle terminals.