3D Semantic Occupancy Prediction for Unmanned Tracked Vehicles in Complex Off-Road Environments

A novel three-dimensional (3D) semantic occupancy prediction method was proposed to analyze and handle the complex off-road environments characterized with diverse geometric, terrain, and road surface features. Firstly, a 3D semantic label was achieved based on a unified framework with the integration of image and LiDAR data. And then, the 3D semantic and occupancy labels were enriched with the Bayesian densification algorithm for the sparse point clouds in off-road scene. Finally, a 3D semantic occupancy grid map was generated, incorporating the size, position, and semantic attributes of environment objects. Experiment results show that the proposed method can extract and represent effectively the 3D environmental information in challenging off-road scenarios, providing a robust foundation for enhanced planning and decision-making in unmanned tracked vehicles.