Hierarchical-Controlled Robot Visual Navigation System Using Line Segment-based Perspective Mapping

Robot navigation systems in unstructured environments are increasingly common but limited by the price of sensors and computational devices. Most robot navigation systems depend on the high-price multi-line lidar to perceive the environment via constructing a 3D grid map directly. The low-price RGB cameras are usually neglected because reconstructing the 3D environment from images needs lots of computational resources. To solve this problem, a hierarchical-controlled robot visual navigation system (HCRVNS) using the RGB camera only to perceive the environment with a low computational burden is proposed in this paper. In HCRVNS, edge points of the drivable area are first extracted by lite semantic segmentation and then transformed into the ground via a transformation matrix generated by an optimization model. Compared to the wide adoption of grid-based maps, the more effective line segment-based map is generated from the edge points on the ground via the proposed adaptive clusters algorithm. An optimization-based path planning algorithm is proposed to balance the total length, curve, and distance to obstacles to match the line segment-based map. Compared to others using sequence structure, a hierarchical environment builder-controller structure with different frequencies is constructed as the backbone of HCRVNS to guarantee its real-time performance. Real-world experiments validate the robustness and effectiveness of HCRVNS, showing the proposed system has similar performance compared to its multi-line lidar-based counterparts.