Probabilistic traversability map generation using 3D-LIDAR and camera

Estimating the traversability of rough terrain is a critical task for an outdoor mobile robot. While classifying structured environment can be learned from large number of training data, it is an extremely difficult task to learn and estimate traversability of unstructured rough terrain. Moreover, in many cases information from a single sensor may not be sufficient for estimating traversability reliably in the absence of artificial landmarks such as lane markings or curbs. Our approach estimates traversability of the terrain and build a 2D probabilistic grid map online using 3D-LIDAR and camera. The combination of LIDAR and camera is favoured in many robotic application because they provide complementary information. Our approach assumes the data captured by these two sensors are independent and build separate traversability maps, each with information captured from one sensor. Traversability estimation with vision sensor autonomously collects training data and update classifier without human intervention as the vehicle traverse the terrain. Traversability estimation with 3D-LIDAR measures the slopes of the ground to predict the traversability. Two independently built probabilistic maps are fused using Bayes' rule to improve the detection performance. This is in contrast with other methods in which each sensor performs different tasks. We have implemented the algorithm on a UGV(Unmanned Ground Vehicle) and tested our approach on a rough terrain to evaluate the detection performance.