Robust positive and negative obstacle detection algorithm in unstructured outdoor environments for unmanned vehicle

In unstructured off-road environments, autonomous driving perception systems must be capable of identifying both positive and negative obstacles. Current methods, such as ground point cloud segmentation algorithms or direct feature extraction from obstacles, are often susceptible to interference from environmental terrain. To address these challenges, this paper proposes a multi-sensor fusion approach. For positive obstacle detection, the Patchwork++ algorithm is robustly enhanced by integrating IMU data and an Adaptive Smoothing Module (ASM). Additionally, dedicated operators are introduced to model mound-like terrains and extract mound point clouds, thereby achieving higher-precision ground point cloud segmentation. For negative obstacle detection, a side-mounted LiDAR configuration is employed to improve the radial resolution of point clouds in front of the vehicle. To enhance the algorithm’s adaptability to various LiDAR types, a geometric model of the LiDAR scanning structure is developed, enabling real-time generation of high-quality simulated ring data through online iteration. This facilitates synchronized detection across heterogeneous LiDAR sensors. Subsequently, geometric features are extracted from the point clouds and integrated with voxel-based Bayesian inference, resulting in a robust negative obstacle detection algorithm. The proposed method was validated using a self-constructed unstructured terrain ground truth dataset and the open-source RELLIS-3D dataset. Experimental results demonstrate that the proposed approach outperforms state-of-the-art methods in both ground point cloud segmentation and negative obstacle point cloud extraction. Finally, the integration of these two algorithms yields a robust positive and negative obstacle perception system for off-road environments, providing reliable support for subsequent decision-making, planning, and control tasks.