A Two-Stage Clustering Method for Point Clouds Based on Cooperative Perception of Vehicle-to-Vehicle
Shuai Han, Huawei Liang, Zhiyuan Li · 2024
The detection of occluded objects and small objects (including distant objects) by autonomous vehicles is still at a low level. Even multi-sensor fusion cannot solve this problem well. V2V cooperative perception has over-the-horizon and wide-range perception capabilities, and can effectively perceive occluded objects and small objects. It has a vast potential for development and a very high performance ceiling. In this paper, a V2V cooperative perception framework is established. We partition the sensing region and perform different information fusion and clustering strategies in different regions to achieve point-cloud level information fusion with low information transmission cost. We also specifically optimize the positioning error and information fusion error of the cooperative vehicles. We devise a two-stage clustering algorithm for V2V cooperative perception. In the first stage, we propose Adaptive Grid Refinement Clustering (AGRC), which achieves better preliminary clustering results in a very short time. In the second stage, we design Scan Line Run Clustering for Multi-vehicle Collaboration (SLR-MVC) to segment the under-segmented objects further, while correcting the object boundaries and eliminating the effects of noise points and point clouds missed by ground segmentation to get accurate clustering results.