A Terrain Segmentation Method Based on U-MobileNet for Navigation of Outdoor Robots

Terrain segmentation is widely used in the recognition and navigation of outdoor mobile robots, in which deep learning-based segmentation methods have achieved remarkable performance. However, most of the segmentation networks can hardly meet the real-time requirement for outdoor robots. A novel terrain segmentation approach is proposed for segmentation terrains in the navigation of robots based on U-MobileNet. U-MobileNet, a lightweight segmentation network, is formed by transposing the feature extraction structure of MobileNetV2 to the contracting path of U-Net. The network efficiently reduce the computational cost and the error in segmentation by applying depthwise separable convolution, inverted residual and linear bottleneck. The experimental results show that the proposed lightweight neural network has good real-time performance with high-accuracy.