Scholarlink

<p>The intelligent vehicle-mounted unmanned aerial vehicle (UAV) system supports functions such as autonomous return and smart track, which can meet the needs of outdoor exploration, navigation planning, and assisted driving in complex scenarios. Due to the characteristics of mutual occlusion, sparse feature distribution, and complex background of traffic objects from the perspective of UAVs, object detection algorithms have low detection accuracy for small objects and are prone to missed and false detections. In order to improve the detection accuracy of object detection algorithms while ensuring real-time inference, the object detection algorithm based on you only look once version 8&#xa0;s (YOLOv8s), namely remote sensing for autonomous driving (RSAD)-YOLO, is proposed. Firstly, the group no-local attention-cross-stage partial bottleneck with two convolutions (GNA-C2f) and the dual branch-spatial pyramid pooling (DB-SPPF) modules were designed to enhance and fuse multi-scale features. Then, a loss function was proposed to accelerate the convergence speed of the model and improve the detection accuracy. Finally, the structure of the YOLOv8s network model was improved by proposing a cross-level fusion mechanism, and the detection head was redesigned to make object localization more accurate. We set YOLOv8s as the baseline algorithm in this study. The experimental results show that on the RSAD dataset, compared to the baseline algorithm, the mAP@0.5 evaluation indicator of the RSAD-YOLO algorithm has increased by 11.8%. Meanwhile, the reasoning speed reaches 137.6 frames per second (FPS), and the parameter quantity is 12.2 million (M), which basically remains unchanged compared with the baseline algorithm. Compared with the state-of-the-art (SOTA) algorithms, the RSAD-YOLO object detection method improves detection accuracy while maintaining the real-time performance of the baseline algorithm. It has stronger generalization ability and effectiveness and is suitable for object detection tasks based on images captured by UAVs</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Advancing traffic object detection in complex environments: a deep learning object detection approach with vehicle-mounted UAV data for traffic scene perception

  • 期刊论文

  • 发表日期：

    2025年12月21日

查看全文
  • Yang Liu,
  • Hongyu Sun,
  • Weiqin Li,
  • Yuyang He

摘要

The intelligent vehicle-mounted unmanned aerial vehicle (UAV) system supports functions such as autonomous return and smart track, which can meet the needs of outdoor exploration, navigation planning, and assisted driving in complex scenarios. Due to the characteristics of mutual occlusion, sparse feature distribution, and complex background of traffic objects from the perspective of UAVs, object detection algorithms have low detection accuracy for small objects and are prone to missed and false detections. In order to improve the detection accuracy of object detection algorithms while ensuring real-time inference, the object detection algorithm based on you only look once version 8 s (YOLOv8s), namely remote sensing for autonomous driving (RSAD)-YOLO, is proposed. Firstly, the group no-local attention-cross-stage partial bottleneck with two convolutions (GNA-C2f) and the dual branch-spatial pyramid pooling (DB-SPPF) modules were designed to enhance and fuse multi-scale features. Then, a loss function was proposed to accelerate the convergence speed of the model and improve the detection accuracy. Finally, the structure of the YOLOv8s network model was improved by proposing a cross-level fusion mechanism, and the detection head was redesigned to make object localization more accurate. We set YOLOv8s as the baseline algorithm in this study. The experimental results show that on the RSAD dataset, compared to the baseline algorithm, the mAP@0.5 evaluation indicator of the RSAD-YOLO algorithm has increased by 11.8%. Meanwhile, the reasoning speed reaches 137.6 frames per second (FPS), and the parameter quantity is 12.2 million (M), which basically remains unchanged compared with the baseline algorithm. Compared with the state-of-the-art (SOTA) algorithms, the RSAD-YOLO object detection method improves detection accuracy while maintaining the real-time performance of the baseline algorithm. It has stronger generalization ability and effectiveness and is suitable for object detection tasks based on images captured by UAVs