一种两阶段路面三维点云目标检测算法

A Two-Stage 3D Point Cloud Object Detection Algorithm for Road Surfaces

针对基于深度学习的三维点云目标检测算法在更换场景或设备的情况下容易出现网络性能无法维持、可迁移性较差的问题,提出一种准确、灵活且迁移性较好的两阶段三维点云目标检测算法(AF3D):在第一阶段,对采集到的激光点云使用分段拟合算法去除路面,并使用具有噪声应用的基于密度的空间聚类(DBSCAN)算法对非地面点云进行聚类,得到若干个聚类簇;在第二阶段,搭建点云全连接网络(PFC-Net)对聚类簇提取特征并进行分类。试验结果表明:该算法在KITTI数据集上可实现良好的检测性能,且在实车数据集上对汽车、行人、骑行者的检测精度分别为69.74%,41.25%、54.33%,具有较好的可迁移性。

The 3D point cloud object detection algorithm based on deep learning is prone to issues such as inability to maintain network performance and poor transferability when changing scenes or devices.To address this issue,this article proposes an Accurate,Flexible,and highly transferable two-stage 3D point cloud object detection algorithm(AF3D).In the first stage of the AF3D detection algorithm,a segmented fitting algorithm is used to remove the road surface from the collected laser point cloud,then DBSCAN algorithm is used to cluster non-ground point clouds and obtain several clustering clusters.In the second stage of the AF3D detection algorithm,a point cloud fully connected network PFC-Net is established,and features are extracted and classified.Through experiments,it has been proven that this algorithm can achieve good detection performance on public KITTI datasets,and the detection accuracy for cars,pedestrians,and cyclists on real vehicle datasets is 69.74%,41.25%,and 54.33%,respectively,indicating good transferability.