Research on Off-road Environment Perception Based on Multi-sensor Fusion

In response to the challenges of complex and changing terrain, lack of obvious road boundaries, and serious dust situations during vehicle operation in off-road environments, this paper proposes an off-road environment perception architecture that includes ground point segmentation and object detection. Firstly, a non-data-driven approach is studied for ground point segmentation in off-road winding paths to ensure real-time performance and interpretability of the algorithm. Secondly, data features are analyzed to develop a noise point filtering scheme based on point cloud intensity and density features. This helps to remove unwanted noise points in the data. Additionally, considering the characteristics of semi-solid-state LIDAR, an improved Euclidean clustering method is researched for obstacle detection. Finally, the clustering results are fused with the 2D object detection results to obtain the final perception output. The method is validated on data collected in off-road environments, and the results demonstrate that it can effectively distinguish ground points from non-ground points, remove over 97% of dust points, and accurately identify objects.