A Hybrid YOLO-NAS Technology in Transforming Autonomous Vehicle Manufacturing

Abstract

Object detection, the core fundamental task of the environmental perception module for autonomous driving, still faces a bottleneck that is difficult to break through. Four types of interference, namely severe weather, low illumination, partial occlusion, and sensor noise, all reduce detection accuracy, which in turn weakens the reliability of core decision-making for safe driving. This study proposes a hybrid perception scheme based on YOLO-NAS that meets the indispensable requirement of real-time performance. By integrating the NAS-optimized architecture and connecting multi-step processing in sequence, the scheme improves the robustness of recognition in complex scenarios. The proposed framework was evaluated on the nuScenes autonomous driving benchmark using the official dataset split. Experimental evaluation under daytime, nighttime, and adverse weather conditions demonstrates that the proposed approach consistently achieves detection accuracies of 94–96%, outperforming YOLOv8 (84–89%) and YOLOv7 (79–81%). Overall, YOLO-NAS provides performance gains of approximately 5–10% over YOLOv8 and 13–25% over YOLOv7 while maintaining stable inference and improved robustness against partially occluded objects and challenging visual conditions. These results demonstrate that incorporating NAS into the YOLO detection pipeline improves perception reliability and detection consistency, offering a practical solution for enhancing the safety, robustness, and real-world deployment of autonomous vehicle perception systems.