Scholarlink

<p>In modern industrial logistics, the efficient sorting of box-shaped objects is a critical link to ensure the smooth operation of supply chains. However, there remains a lack of corresponding classification and grasping methods for complex scenarios. With the advancement of depth camera technology, obtaining texture and depth information of objects has become increasingly convenient, providing new opportunities to enhance the intelligence level of sorting systems. To this end, this paper proposes an innovative box-shaped object sorting system based on multi-modal information. This method integrates image instance segmentation, template matching algorithms, and robotic arm operations, achieving efficient and accurate sorting of box-shaped objects with varying sizes, colors, and materials. This paper proposes a multi-modal multi-level fusion instance segmentation network, MSAF-YOLO, based on YOLO, which takes RGBD inputs. It is trained using the synthetic dataset Snack-Box. Through multiple sets of experiments, the feasibility of this method was confirmed and the optimal network architecture was determined. Experimental results show that MSAF-YOLO (with 39.320M parameters and 161.336 GFLOPs) shows improvement in the metrics mAP50 and mAP50-95 by 2.6% and 3.99%, respectively, compared to YOLOv8m (RGB). MSAF-YOLO outperforms the state-of-the-art RGBD methods by 1.3% and 1.79% in the same metrics. In nearly 1,000 scene experiments, MSAF-YOLO achieved a successful grasping rate of 92.72%. Our project is available at: <a href="https://github.com/MSAF-YOLO/MSAF-YOLO">https://github.com/MSAF-YOLO/MSAF-YOLO</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

Sorting of box-shaped objects based on multi-modal information

  • 期刊论文

  • 发表日期：

    2025年6月14日

查看全文
  • Yuqiu Wu,
  • Qinmei Chen,
  • Jinlong Shi,
  • Haoran Deng,
  • Suqin Bai

摘要

In modern industrial logistics, the efficient sorting of box-shaped objects is a critical link to ensure the smooth operation of supply chains. However, there remains a lack of corresponding classification and grasping methods for complex scenarios. With the advancement of depth camera technology, obtaining texture and depth information of objects has become increasingly convenient, providing new opportunities to enhance the intelligence level of sorting systems. To this end, this paper proposes an innovative box-shaped object sorting system based on multi-modal information. This method integrates image instance segmentation, template matching algorithms, and robotic arm operations, achieving efficient and accurate sorting of box-shaped objects with varying sizes, colors, and materials. This paper proposes a multi-modal multi-level fusion instance segmentation network, MSAF-YOLO, based on YOLO, which takes RGBD inputs. It is trained using the synthetic dataset Snack-Box. Through multiple sets of experiments, the feasibility of this method was confirmed and the optimal network architecture was determined. Experimental results show that MSAF-YOLO (with 39.320M parameters and 161.336 GFLOPs) shows improvement in the metrics mAP50 and mAP50-95 by 2.6% and 3.99%, respectively, compared to YOLOv8m (RGB). MSAF-YOLO outperforms the state-of-the-art RGBD methods by 1.3% and 1.79% in the same metrics. In nearly 1,000 scene experiments, MSAF-YOLO achieved a successful grasping rate of 92.72%. Our project is available at: https://github.com/MSAF-YOLO/MSAF-YOLO.