RE-YOLO: Vehicle Detector Based on Knowledge Distillation YOLOv8 for Autonomous Driving Scenarios

Vehicle detection tasks in autonomous driving scenarios face challenges such as illumination changes, motion blur, and target occlusion. This paper proposes a vehicle detector RepViT efficient-you only look once (RE-YOLO) for autonomous driving scenarios based on knowledge distillation you only look once Version 8 (YOLOv8), which targets specific requirements in the field of autonomous driving, such as real-time, high precision, low resource consumption, and adaptability to complex scenarios. First, RepViT Block in the newly proposed RepViT was used to replace Cross Stage Partial with 2 fusion (C2f) in the backbone, and a new C2f-RVB (RepViT Block) module was designed. The computational parameters, floating point operations (FLOPs), and model size were reduced by 19.6%, 13.6%, and 18.3%, respectively, and the running speed was increased by 1.9%. Then, the efficient multiscale attention (EMA) mechanism was integrated with C2f-RVB, and a new high-level dynamic sampling module was independently developed to replace the neck part. Afterward, the original detection head, Detect, was replaced with the new lightweight detection head, Efficient Head, designed in this paper to further the accuracy and efficiency of vehicle detection. Finally, knowledge distillation was performed on the improved model, and the model was called RE-YOLO. The experimental results show that compared with the baseline YOLOv8, RE-YOLO reduces computational parameters by 47.8%, FLOPs by 38.3%, and model size by 44.7%. In addition, RE-YOLO’s mean average precision) increases by 1.2%, and its running speed increases by 28.1%, reaching 80.6 frames per second. These results demonstrate that RE-YOLO offers substantial improvements in detection performance and processing speed, providing robust support for autonomous driving systems’ environmental perception needs and setting a foundation for future advancements in vehicle detection for autonomous applications.