Terrain Environment Estimating Method for off-Road Vehicle Anticipated Driving Area Based on Stereo Vision

<div class="section abstract"><div class="htmlview paragraph">Off-road vehicles are required to traverse a variety of pavement environments, including asphalt roads, dirt roads, sandy terrains, snowy landscapes, rocky paths, brick roads, and gravel roads, over extended periods while maintaining stable motion. Consequently, the precise identification of pavement types, road unevenness, and other environmental information is crucial for intelligent decision-making and planning, as well as for assessing traversability risks in the autonomous driving functions of off-road vehicles. Compared to traditional perception solutions such as LiDAR and monocular cameras, stereo vision offers advantages like a simple structure, wide field of view, and robust spatial perception. However, its accuracy and computational cost in estimating complex off-road terrain environments still require further optimization. To address this challenge, this paper proposes a terrain environment estimating method for off-road vehicle anticipated driving area based on stereo vision. First, a real-vehicle data acquisition platform was constructed using data converters such as Kvaser and stereo vision cameras. Second, by integrating stereo vision perception boundary constraints, vehicle body axis expansion techniques, and the Constant Turn Rate and Velocity (CTRV) model, the method reliably extracts environmental information from the intended driving region, which could avoided the use of global information estimation and reducing computational costs. Building on this foundation, we developed a data delay timestamp synchronization mechanism and a circular buffer to enable the acquisition of vehicle dynamics signals, as well as road feature categories and point cloud data for 16 types of roads, including dirt roads, highways, rocky roads, and others. Thirdly, by establishing U-Net semantic segmentation model, Mean Intersection over Union and Crossentropy loss, pavement type identification is completed. Finally, road unevenness estimation was performed using a Random Sample Consensus (RANSAC) algorithm optimized with a Voxel Grid Filter. Experimental results demonstrate that the proposed estimation method enables off-road vehicles to accurately and robustly estimate environmental information such as road type and road unevenness with relatively low computational costs, achieving an average Intersection over Union of 85.96 and a loss function as low as 0.38.</div></div>