Stair-Pothole Detection and Distance Estimation for Unmanned Robots

Nowadays, the application scenarios for autonomous robots are increasingly expanding. Autonomous robots decide whether to navigate through static obstacles in the environment based on their exploration of indoor and outdoor surroundings and their own mobility capabilities. Among these challenges, detecting obstacles such as stairs (protruding obstacles) and potholes (depressed obstacles) is particularly difficult. However, existing datasets mostly focus on structured roads like urban streets, lacking data on indoor and outdoor environments with irregular obstacles. We propose a stair-pothole dataset to fill this gap. This is an instance segmentation dataset containing 8,451 images and over 14,000 labels. We proposes a method for instance segmentation and ranging of protruding and depressed obstacles based on an improved YOLOv8 model. First, the CBAM attention module is added to enhance the detection accuracy of small objects. Second, the WIoU loss function is used to overcome the limitations of the default CIoU loss function in YOLOv8. Finally, the distance of stair and pothole obstacles is estimated using binocular ranging. Validation of the improved model shows that the experimental results achieve a 5.4<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> point increase in mAP50-95 compared to the original YOLOv8n-seg model, with a detection speed of 167 fps. The improved model has been deployed on the test platform, achieving a detection speed of 112 fps and an mAP of 84.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, with distance measurement error less than 10 cm. The experimental results indicate that the improved model can provide semantic recognition and ranging of protruding and depressed obstacles in indoor and outdoor environments for researchers in related fields.