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

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.