Negative obstacle detection and avoidance using YOLOv8 and depth profile analysis for autonomous mobile navigation

Autonomous ground robots must effectively detect and avoid negative obstacles such as holes and ditches. Existing approaches relied on single sensors such as LiDAR and cameras, but each of these is insufficient in most cases. Recent improvements have involved the use of machine learning and the fusion of multiple sensors to account for the limitations of single sensors. However, challenges due to computational cost, fusion complexity, and efficiency still exist for these methods. To minimize the complexity involved in fusing multiple sensors this paper presents a negative obstacle avoidance system that integrates a YOLOv8n model with Intel RealSense depth camera data and a 2D laser scan for negative obstacle detection within the ROS2 navigation system. The proposed system combines YOLOv8-based detection with depth validation to reduce false detections. The resulting detection is represented as a laser scan and fused with positive obstacle laser scan data, enabling easy compatibility with the existing ROS2 navigation stack. The system was tested in a simulation environment and a custom dataset was created and used to train the YOLOv8 model for the system. The simulation results of our proposed system demonstrate reliable and effective detection and avoidance while ensuring minimal computational cost and complexity.