| 摘 要: 针对海事场景下2D目标检测缺乏深度信息、激光雷达3D检测对稀疏及遮挡目标精度受限的问题，提出一种改进FPointNet的3D目标检测算法。通过深度可分离卷积强化点云几何特征提取，在T-Net中引入3DCAM提升目标关键区域感知能力，在边界框回归网络中引入适配海事场景的CSDF BLOCK实现多尺度特征融合。实验基于仿真与真实场景的混合数据集开展，结果显示，改进的算法在mAP@0.5和mAP@0.7指标分别达到84.54%和51.05%，较基线模型分别提升4.41%和3.32%，在特定条件下改善了海事场景稀疏、遮挡目标检测精度不足的问题，验证了改进策略的有效性。 |
| 关键词: 三维目标检测 注意力机制 多尺度融合 船舶检测 |
|
中图分类号:
文献标识码:
|
| 基金项目: 江苏省重点研发计划项目(BE20222136)；闽东水产品精深加工福建省高校工程研究中心开放基金(NSHY2025Y03) |
|
| Maritime Scene 3D Object Detection Algorithm Based on Improved FPointNet |
|
LiuYiJiang1, SU Zhen2, HAN Kunhuang3, Wang wei2
|
1.Jiangsu University of Science and Technology School of Computer, ;2.Marine Equipment and Technology Institute, Jiangsu University of Science and Technology;3.Engineering Research Center of Mindong Aquatic Product Deep-Processing, Fujian Province University
|
| Abstract: To address the problems that 2D object detection in maritime scenarios lacks depth information and LiDAR-based 3D detection suffers from limited accuracy for sparse and occluded targets, an improved FPointNet-based 3D object detection algorithm is proposed. Depthwise Separable Convolution is adopted to enhance the geometric feature extraction of point clouds; the 3DCAM is incorporated into the T-Net to improve the model's perception capability for key target regions; and the maritime scenario-adapted CSDF BLOCK is introduced into the bounding box regression network to achieve multi-scale feature fusion. Experiments are conducted on a mixed dataset combining simulated and real-world scenarios. The results show that the improved algorithm achieves 84.54% and 51.05% in the metrics of mAP@0.5 and mAP@0.7 respectively, with respective improvements of 4.41% and 3.32% over the baseline model. It alleviates the issue of insufficient detection accuracy for sparse and occluded targets in maritime scenarios under specific conditions, verifying the effectiveness of the proposed improvement strategies. |
| Keywords: 3D Object Detection Attention Mechanism Multi-scale Fusion Ship Detection |
