Long-Range LiDAR Vehicle Detection Through Clustering and Classification for Autonomous Racing

원문 초록

With the expansion of autonomous driving technology, autonomous racing has been actively studied in recent years. For safe autonomous racing, fast computation speed and wide detection range are essential. However, existing methods can perform detection only in short range due to the sparsity of LiDAR data. While some methods have been proposed to address this limitation, they often neglect real-time processing, making them unsuitable for racing environments. To address this, we propose a novel clustering and classification-based method specifically designed to enhance long-range vehicle detection while maintaining computational efficiency. First, our method employs a lightweight road segmentation and ground removal module to eliminate irrelevant data, significantly reducing computational overhead without down-sampling. Second, we introduce a 2D Bird’s Eye View (BEV)-based clustering approach, which is faster and more robust than traditional 3D Euclidean clustering, to generate object candidates efficiently. Additionally, a machine learning classifier trained on long-range vehicle features is incorporated to enhance detection accuracy at extended distances. Finally, range-view-based verification is performed to refine and finalize detection results, ensuring high reliability. The proposed method is evaluated on the aiMotive dataset, long-range dataset. Experimental results demonstrate that our approach achieves over 80% detection accuracy at distances exceeding 50 meters, with a computation time of only 25 ms per frame. These results underscore the novelty and effectiveness of the proposed method as an optimized solution for long-range vehicle detection in real-time autonomous racing environments.