基于PointNet++的室外场景三维点云多目标检测方法

Multi-Object Detection Method of 3D Point Cloud in Outdoor Scene Based on PointNet++

针对室外场景三维点云的稀疏性对目标检测带来的挑战问题,设计一个基于PointNet++的点云检测方法。该方法首先预处理点云,获取感兴趣区域点云;再聚类点云,对物体进行分割;接着通过PointNet++检测,获得目标的类别结果;最后通过三维boundingbox获得目标物体的长宽高及朝向。为验证该方法的有效性,用16线velodyne激光雷达采集室外真实场景的数据,并制作样本集进行网络训练。最终结果验证,该方法能获得较高的检测准确率,并满足实时性要求。

The sparsity of three-dimensional point clouds in outdoor scenes poses a challenge to multi-object detection.To solve this problem,a point cloud detection method based on PointNet++is designed.Firstly,the point cloud is preprocessed to get the point cloud in the region of interest;secondly,the point cloud is clustered to segment the object;thirdly,the result of segmentation is detected by PointNet++to get the category result of the object;finally,the length,width,height and orientation of the object are obtained by the design of three-dimensional bounding box generation method.In order to verify the validity of the framework,we use 16-line velodyne lidar to collect outdoor real scene data,and make sample sets for network training.The final results show that our method can achieve high detection accuracy and meet the real-time requirements.