Visual Road Boundary Detection based Local Path Planner

Abstract

We propose a real-time visual road detection based local path planner for autonomous navigation in unstructured environments. The system consists of two parts: road detection and local path planning. Road boundary model based on convolution neural network is trained for local road detection in image frame. After perspective transformation, the result is transformed to the vehicle’s frame for path planning. The road midline generated from the road boundaries is used to represent the reference local path. Priori information with the global navigation topology map is combined to solve multi-roads selection. Finally, data fusion with historical detection results and predictive control of vehicle model is used for path optimization. In addition, obstacle avoidance combined with LIDAR perception is also taken into account to complete an autonomous navigation task in complex scenarios.A large number of experiments with various road conditions shows that the proposed method has a good adaptability and versatility for unstructured scenarios. The algorithm can effectively reduce the dependence on absolute positioning and LIDAR perception. So it is of great importance for autonomous navigation in scenarios without satellite or in unknown scenarios.