基于点云多帧融合的相邻障碍物聚类分割方法

Segmentation method of adjacent obstacles based on multi-frame fusion point clouds

相邻障碍物的分割是无人驾驶领域的技术难点,低线激光雷达点云稀疏,无法聚类远距离物体,但激光雷达线束越多越昂贵。为了实现低成本聚类分割相邻障碍物,实验场景选取常用交通场景对象相邻的人/人、人/车,提出了一种基于多帧融合的相邻障碍物分割方法。基于惯性测量单元、激光雷达融合多帧点云,解决了低线激光雷达因分辨率低而无法聚类远距离相邻行人的问题。提出改进的欧式聚类,加入自适应阈值和向量角度约束两个新的分割标准,提高相邻障碍物的分割效果。实验结果表明,该方法具有成本低、聚类精准等特点,与单帧传统欧式聚类算法相比,该方法针对相邻障碍物分割的准确度提升约30.7%,对低线激光雷达在障碍物聚类以及后续的检测具有一定参考意义。

The segmentation of adjacent obstacles is a technical difficulty in the field of driverless vehicles.The low-line light detection and ranging(LIDAR)point clouds are sparse,and they can't cluster long-distance objects.However,the more LIDAR wire beams,the more expensive.In order to realize low-cost clustering segmentation of adjacent obstacles,the adjacent people and vehicles of common traffic scene objects were selected as the experimental scenes,and a adjacent-obstacle segmentation method based on multi-frame fusion was proposed.Based on the inertial measurement unit(IMU)and LIDAR fusion of multi-frame point cloud,it could solve the problem that low-line LIDAR is unable to cluster long-distance adjacent pedestrians due to its low resolution.The improved Euclidean clustering was proposed,and two new segmentation criteria of adaptive threshold and vector angle constraint were applied into this algorithm to improve the segmentation effect of adjacent obstacles.The experimental results show that this method has the characteristics of low cost and accurate clustering.Compared with the single-frame Euclidean cluster,the accuracy of this method for segmenting adjacent obstacles improves by about 30.7%,which has certain reference significance for the clustering and detection of low-line LIDAR in obstacles.