Scholarlink

<p>In the practical inland-river freight scenarios, the ship point cloud encounters challenges such as uneven density, large-scale variations, and a tendency to lose structural information. To address these issues, a novel ship point cloud target detection method is proposed. The method adopts the point-voxel region-convolutional neural network (PV-RCNN) as the baseline and enhances it from three perspectives. Specifically, a density-weighted voxel-farthest point sampling method is proposed first for key point sampling. Subsequently, a coordinate-mixed attention mechanism is developed for feature extraction. Moreover, a dynamic graph convolution feature aggregation module based on the candidate bounding box is proposed, enabling adaptation to the characteristics of point clouds in the inland-river freight scenario. This work is validated using a ship point cloud dataset from a practical project deployed at BaiXian Lake in Zhouzhuang. Compared with the baseline network, the average precision of the LiDAR side-scan ship point cloud and the BEV-scan ship point cloud have been enhanced by 2.3% and 3.8%, respectively. Experimental results indicate that the proposed method compensates for the limitations of the baseline PV-RCNN in terms of extracting long-range and structural features of the target. Consequently, it boosts the precision and robustness in target detection of ship point clouds within the inland-river freight scenario.</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

Dynamic graph convolution and mixed attention mechanism based ship point cloud target detection

  • 期刊论文

  • 发表日期：

    2025年2月4日

查看全文
  • Yi Zhou,
  • Wenkai Zhang,
  • Yuwei Min,
  • Jianfeng Yang,
  • Tianqi Yu

摘要

In the practical inland-river freight scenarios, the ship point cloud encounters challenges such as uneven density, large-scale variations, and a tendency to lose structural information. To address these issues, a novel ship point cloud target detection method is proposed. The method adopts the point-voxel region-convolutional neural network (PV-RCNN) as the baseline and enhances it from three perspectives. Specifically, a density-weighted voxel-farthest point sampling method is proposed first for key point sampling. Subsequently, a coordinate-mixed attention mechanism is developed for feature extraction. Moreover, a dynamic graph convolution feature aggregation module based on the candidate bounding box is proposed, enabling adaptation to the characteristics of point clouds in the inland-river freight scenario. This work is validated using a ship point cloud dataset from a practical project deployed at BaiXian Lake in Zhouzhuang. Compared with the baseline network, the average precision of the LiDAR side-scan ship point cloud and the BEV-scan ship point cloud have been enhanced by 2.3% and 3.8%, respectively. Experimental results indicate that the proposed method compensates for the limitations of the baseline PV-RCNN in terms of extracting long-range and structural features of the target. Consequently, it boosts the precision and robustness in target detection of ship point clouds within the inland-river freight scenario.