Obstacle Detection by Autonomous Vehicles: An Adaptive Neighborhood Search Radius Clustering Approach

For autonomous vehicles, obstacle detection results using 3D lidar are in the form of point clouds, and unevenly distributed space. Clustering is a common means for cloud processing; however, improper selection clustering thresholds can lead to under-segmentation or over-segmentation resulting false missed obstacles. In order solve these problems, new method was required. Firstly, we applied distance-based filter ground segmentation algorithm, pre-process original cloud. Secondly, proposed an adaptive neighborhood search radius based on analysis relationship between spatial distribution, adopting pitch angle horizontal resolution lidar, determine threshold. Finally, vehicle platform offline driving KITTI dataset were used conduct multi-scene comparative experiments Euclidean method. The real experimental showed that our improved accuracy by 6.94%, F1 score increased 0.0629.