Boundary Recognition for Lawn Mowing Robot Based on Image Segmentation

Accurately recognition of lawn boundaries is the basis for the reliable operation of lawn mowing robots. Current lawn boundary recognition methods rely on pre-buried cables or mapping boundaries through RTK-GPS positioning technology. Both methods are susceptible to positioning errors and environmental changes. Real-time recognition of lawn boundaries based on image segmentation can form a real-time closed loop between path planning and boundary recognition of lawn mowing robots, which improvs the robustness and reliability of robot work. The U-Net network is a simple image segmentation model that is suitable for robots with limited computing resources. However, the results of the binary segmentation of the lawn are often open boundary lines, which are different from the results with some multi-closed cells of the U-Net model in medical images. Therefore, it is difficult to directly apply the U-Net model for accurately lawn segmentation. Considering the characteristics of lawn images and the limited computing resources, this paper introduces an improved U-Net model with a channel-space attention mechanism and a changed loss function, which better solves the problem of lawn boundary recognition. The mdice value of the improved model is 97.7%, which is about 2% higher than the original U-net model.