Abstract

The ability to perceive traversable regions is a crucial prerequisite for truly autonomous mobile systems. In our previous work [8], we introduced P-SVM, a 3D LiDAR-based traversability analysis system that achieves real-time performance in CPU. However, by design P-SVM does not capture features other than geometric features, limiting its performance and generalizability to complex environments. In this work, we introduce an augmented point cloud descriptor that further improves the performance of P-SVM. By exploiting additional features extracted from remission and height information in the point cloud, our model is able to adapt robustly to strong scene changes. Additionally, in the proposed descriptor, remission radial distribution features are introduced to capture local information around keypoints. This addresses the limitation of P-SVM, which focuses only on global features within the cell. Our new P-SVM2 model demonstrates performance almost on par with state-of-the-art deep learning-based methods on the challenging SemanticKITTI [1] dataset in the traversability analysis task. Notably, P-SVM2 is real-time and relies solely on mobile-level CPUs. Moreover, surprising results have been obtained in robustness and generalizability experiments.