A Terrain Recognition Method Based on Semantic Segmentation for Field Robot Under Sample Imbalance

Abstract

Vision-based terrain recognition technology can help the robot understand the surrounding environment and then adjust the control and motion strategy in time. However, most of the existing work focuses on urban environments, and little attention is paid to complex off-road environments. Moreover, the class imbalance in the off-road environment is highly dependent on the specific environment so that the class imbalance problem is more serious, which makes the semantic segmentation of rare classes more challenging. In this study, we propose a visual terrain recognition method based on different weight loss functions for ground mobile robots in the off-road environment with class imbalance. The visual information is input into the improved network for terrain recognition. In order to better identify the rare classes, different weights are assigned to different classes in the loss function of the improved network. After a series of experiments, the results show that our improvement achieves better results than the original model in the off-road environment.