An Autonomous Navigation Approach based on Bird’s-Eye View Semantic Maps

With the continuous expansion of the use of mobile robots, providing them with autonomous navigation capabilities for different environments has become a very active research topic. In most cases, navigation systems are built around sensors like LiDAR which are expensive, not only due to the sensor cost but also to the computational power required for the processing of Point-Clouds. To provide an alternative to such systems, we propose a navigation approach that only requires a front-facing RGB Camera. In this system, every image is processed online using a semantic segmentation model to build a Bird’s-Eye view semantic map, from which a local path and the corresponding motion commands can be calculated. Our method is evaluated first in simulation and later on, in a real mobile robot. The results show that our system enables the robot’s successful navigation and collision avoidance through simulated and real indoor and outdoor environments.