Abstract:A lightweight road-assistant detection algorithm, EBD-YOLO, based on YOLOv5s is proposed to address the problems of high model complexity, computation cost, and difficulty in deployment on resource-limited embedded terminals in existing assisted driving detection algorithms. First, lightweight Transformer model EfficientViT was used as the backbone feature extraction network of YOLOv5s model to reduce network parameters and calculation costs. Secondly, a Focal-GIoU Loss function is proposed for bounding box regression to accelerate model conver-gence and reduce loss. Thirdly, the feature pyramid structure is improved to a weighted bi-directional feature pyramid network (BiFPN) to enhance localization and semantic features. Then, a dynamic head framework is added to unify the attention mechanism with the object detection head to improve its performance. Finally, a Soft-CIoU_NMS algorithm is proposed in the post-processing stage to enhance occluded targets' localization and detection ability and reduce the missed detection rate. We conducted experiments on the KITTI and BDD100K datasets for autonomous driving, and the results showed that the EBD-YOLO model reduced in size by 38.4% and 37.2%, respectively. In comparison, the computational cost was reduced by 48.1%. As measured by mAP@0.5, the detection accuracy improved by 0.5% and 5.8%, respectively, and mAP@0.5:0.95 improved by 2.8% and 7%, re-spectively. These improvements satisfied the requirements for deployment on embedded terminals in cars.