基于特征流的点云目标检测方法

PDF下载HTML

Title:

Feature flow-based point cloud object detection method

作者:

陆军, 邹康成, 李杨

哈尔滨工程大学 智能科学与工程学院, 黑龙江 哈尔滨 150001

Author(s):

LU Jun, ZOU Kangcheng, LI Yang

College of Intelligent Science and Engineering, Harbin Engineering University, Harbin 150001, China

关键词:

激光雷达点云; 目标检测; 特征流; 特征对齐; 时序特征融合; 可变形注意力机制; 鸟瞰视角表示; 多帧点云融合

Keywords:

lidar point cloud; object detection; feature flow; feature alignment; temporal feature fusion; deformable attention mechanism; bird’s-eye view; multi-frame point cloud fusion

分类号:

TP391

DOI:

10.11992/tis.202503005

摘要:

针对现有激光雷达点云三维目标检测方法因点云稀疏性导致的场景信息缺失与目标漏检问题，本文提出一种基于特征流的单阶段三维目标检测算法，该算法通过多帧时空特征融合与动态对齐机制优化检测性能。首先，构建门控网络驱动的多帧融合框架，利用可变形注意力机制协同时空特征提取模块，实现跨帧特征的动态对齐，抑制未对齐特征融合导致的误检；其次，设计时空特征引导的可变形注意力机制，通过目标运动信息预测特征偏移与权重，提升稀疏点云的特征匹配精度；最后，设计层级式特征流提取模块，结合多尺度特征提取与渐进融合策略，增强场景表征能力。实验结果表明，所提算法在NuScenes验证集上的平均精度均值达到63.73%，较体素基准方法提升4.51%，其中摩托车、自行车等小目标检测精度提升超过14%。消融实验结果表明，多帧互补机制使远距离目标(>50 m)召回率提升16.2%，遮挡场景漏检率降低11.8%。本研究为自动驾驶领域稀疏点云三维检测提供了有效方案。

Abstract:

Aiming at the problem of missing scene information and missing target detection caused by the sparsity of point cloud in the existing 3D target detection method of lidar point cloud, this paper proposes a single-stage 3D target detection algorithm based on feature flow, and the algorithm optimizes the detection performance through multi-frame spatio-temporal feature fusion and dynamic alignment mechanism. Firstly, a multi-frame fusion framework driven by gated network is constructed. The deformable attention mechanism is used to cooperate with the spatio-temporal feature extraction module to realize the dynamic alignment of cross-frame features and suppress the false detection caused by unaligned feature fusion. Secondly, a deformable attention mechanism guided by spatio-temporal features is designed to predict feature offset and weight through target motion information, so as to improve the feature matching accuracy of sparse point clouds. Finally, a hierarchical feature flow extraction module is designed to enhance the scene representation ability by combining multi-scale feature extraction and progressive fusion strategy. Experiments show that the proposed algorithm achieves 63.73% mAP on the NuScenes verification set, which is 4.51% higher than the voxel benchmark method, and the detection accuracy of small targets such as motorcycles and bicycles is improved by more than 14%. Ablation experiments show that the multi-frame complementary mechanism increases the recall rate of long-distance targets (>50 m) by 16.2%, and reduces the missed detection rate of occlusion scenes by 11.8%. This study provides an effective solution for three-dimensional detection of sparse point clouds for autonomous driving.