Obstacle Avoidance Method with Local Semantic Map Generation for Self-Driving Cars

When implementing a control system for a ground-based highly automated vehicle (HAV) used in agriculture, a significant factor is the low cost of the equipment used, so cameras using without lidars for navigation, avoiding obstacles, and ensuring movement safety seems promising. Ensuring the safety of movement of HAV consists of developing methods and tools that allow to avoid static and dynamic obstacles and make an emergency stop if it is impossible to avoid. The main problem when using cameras is the lack of information about the distance to objects. The article describes methods and software tools that allow transforming an image into a top view (birds-eye-view) projection, and use information about free space for movement along with information about obstacles in order to improve the quality of obstacle avoidance algorithms compared to classical methods that use frontal projection. Object detection implemented with YOLOv8 neural network, free space for movement detects with semantic segmentation tool based on MobileNetV3 network architecture, also software tools for combining several data sources to build a local perception map are developed. Linear and angular speed of the HAV based on calculating the optimal direction of movement for avoiding obstacles. Testing of the developed methods and software implemented in the Webots simulation environment. Software developed with ROS2.