Identification of Trees in an Obstacle-rich Forested Environment for Off-Road Autonomous Ground Vehicles

In off-road environments, accurate classification of different obstacle types can significantly improve the path planning decision making for off-road autonomous systems. More specifically, in the context of forested environments, the ability of the scene perception algorithm to correctly distinguish trees from solid obstacles is of special importance. Classification of trees as solid obstacles would unnecessarily limit the traversable area. While tree trunks, in many cases, can be correctly classified as solid obstacles, tree crowns including leaves and thin branches might not be a collision hazard for navigation. From the safe navigation standpoint, the ability of making distinction between safe parts and non-safe parts of a tree can help the scene perception algorithm remain on a reliable level while improving the optimized path. To achieve the goal, the first step is developing a solution to correctly distinguishing trees from solid obstacles. This paper proposes a new framework for such a solution. Relevant features were extracted from the scene data generated by a physics-based simulator called MAVS and different classification models are used to investigate the effectiveness of the extracted features. The results confirmed that the proposed methodology including the extracted features and selected parameters were relevant enough to result in the desired prediction accuracy level to distinguish between obstacle and non-obstacle regions, as well as tree and solid obstacle regions.