Environmental Perception for Autonomous Forestry Vehicles

Abstract

Abstract Autonomous navigation in forestry environments presents significant challenges due to complex, unstructured terrain with varying visibility conditions. This thesis presents a novel sensor fusion approach integrating LiDAR and stereo camera data for enhanced terrain mapping in forestry applications. The project develops an uncertainty-aware fusion framework based on Kalman filtering that effectively combines the high accuracy of LiDAR with the dense coverage of stereo camera data, while properly accounting for each sensor’s unique error characteristics and uncertainties. Additionally, a dynamic voxel-based representation is implemented that adapts map resolution to terrain complexity, optimizing memory usage while maintaining high fidelity in regions of interest. Experimental results demonstrate measurable improvements in various dimensions: the dynamic voxelization reduced memory usage by 31.65% and improved map update time by 44.27% compared to traditional fixed-size voxel grids, while maintaining mapping quality. Testing on real-world autonomous navigation routes showed that the proposed approach enables more complete trajectory following compared to the previous single-sensor approach, achieving path lengths significantly closer to the planned trajectory - for instance, 38.99m compared to 18.99m in one test. This work demonstrates that intelligent fusion of complementary sensors, combined with adaptive mapping techniques, can significantly improve terrain perception for autonomous vehicles operating in challenging off-road environments.