A Novel Terrain Segmentation Approach for Scene Understanding of Field Robots

Terrain segmentation is crucial for field robots' navigation, path planning, and map building in unstructured environments. However, it is still a challenge for most existing vision-based methods to segment terrains with satisfactory accuracy and real-time performance. Deep learning-based approaches have been proven to be quite competitive for image segmentation. However, most previous segmentation networks have deep networks and explosive growth of the number of parameters, which can hardly satisfy the inference speed of onboard computers. Therefore, a novel terrain segmentation method founded on CSPResnet is designed for segmenting terrains of field robots in this study. It fuses some advantages of several state-of-the-art networks and has a novel structure suitable for on-board computers. The proposed method reduces the computational cost for segmenting terrains by cross-stage partial, spatial pyramid pooling, and nearest interpolation. Finally, some comparison and ablation studies were made on the HDU-Terrain dataset. We collected this dataset from a field robot's perspective. It is quite different from the existing benchmark dataset for unmanned driving and has more than 4000 frames of pixel-wise annotation, with many frequently encountered unstructured terrain types. Experimental results prove that the proposed terrain segmentation network named TSCSPnet balances real-time performance with high accuracy and has the potential to be applied to various field robots.