Scholarlink

<p>Object detection, particularly for small and deformed targets in complex backgrounds, remains a significant challenge for service robots. To address this, we propose a novel lightweight YOLO-MRC (You Only Look Once-MobileViT-RepPoints-CTAM) network. First, an improved C3-RepPoints module is proposed for finer localization and classification of objects of varying scales and deformations. Second, a lightweight hybrid architecture is presented for efficient small object detection, combining convolutional neural networks (CNNs) and vision transformers (ViTs). Finally, an improved efficient decoupled head (EDH) is used to enhance detection performance by separating classification and regression tasks. Furthermore, the convolutional triplet attention mechanism (CTAM) is introduced to aggregate crucial semantic information. Experimental results on the ODSR-HIS dataset show that YOLO-MRC reduces model parameters by 21.3% while improving mean average precision (mAP@0.5) by 1.3%, with notable gains for small-scale and deformed objects. YOLO-MRC also demonstrates excellent generalization on PASCAL VOC and COCO datasets, improving mAP@0.5 by 4.4% and 10.7%, respectively, compared to YOLOv5n, all achieved with only 1.4M parameters and 5.5G FLOPs. This underscores YOLO-MRC’s balance between model size, latency, and accuracy, making it well suited for service robots. The code is available at <a href="https://github.com/lvyongshjd/YOLO-MRC">https://github.com/lvyongshjd/YOLO-MRC</a>.</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

An optimized lightweight YOLO-MRC framework for enhanced obstacle detection in service robots

  • 期刊论文

  • 发表日期：

    2025年8月31日

查看全文
  • Junyan Tian,
  • Ning Ma,
  • Keya Huang,
  • Yong Lv,
  • Wenzheng Chi,
  • Lining Sun

摘要

Object detection, particularly for small and deformed targets in complex backgrounds, remains a significant challenge for service robots. To address this, we propose a novel lightweight YOLO-MRC (You Only Look Once-MobileViT-RepPoints-CTAM) network. First, an improved C3-RepPoints module is proposed for finer localization and classification of objects of varying scales and deformations. Second, a lightweight hybrid architecture is presented for efficient small object detection, combining convolutional neural networks (CNNs) and vision transformers (ViTs). Finally, an improved efficient decoupled head (EDH) is used to enhance detection performance by separating classification and regression tasks. Furthermore, the convolutional triplet attention mechanism (CTAM) is introduced to aggregate crucial semantic information. Experimental results on the ODSR-HIS dataset show that YOLO-MRC reduces model parameters by 21.3% while improving mean average precision (mAP@0.5) by 1.3%, with notable gains for small-scale and deformed objects. YOLO-MRC also demonstrates excellent generalization on PASCAL VOC and COCO datasets, improving mAP@0.5 by 4.4% and 10.7%, respectively, compared to YOLOv5n, all achieved with only 1.4M parameters and 5.5G FLOPs. This underscores YOLO-MRC’s balance between model size, latency, and accuracy, making it well suited for service robots. The code is available at https://github.com/lvyongshjd/YOLO-MRC.