Vision-Based Autonomous Navigation with a Probabilistic Occupancy Map on Unstructured Scenarios

Vision-based robotics perception still have a great focus of attention on building systems because of its common availability and low cost. The 3D data produced by the disparity calculation methods in stereo cameras are inaccurate and presents substantial noise. We present here our method to deal with the noisy 3D point cloud produced by stereo camera to build a navigation map and mark obstacles with a probabilistic occupancy map approach. The objective is to integrate continuously the sensor readings marking occupied and free space based on some certainty and accumulate it over time. The output is a navigability map we use to plan a trajectory path. Our main focus is applications like agricultural fields. We have modeled and tested the system fully in simulation and validated it with our real vehicle platform Carina I on unstructured scenarios.