A real-time semantic segmentation method for autonomous driving in surface mine

Semantic segmentation for perception of autonomous vehicles has made great progress with the development of deep learning technique. However, existing related studies mainly focus on urban scenes, thus semantic segmentation for perception of autonomous mining trucks in surface mine remains to be further explored. Considering the importance of dataset in deep learning technique, we first built a semantic segmentation dataset of surface mine. In surface mine environment, it is a fact that the particularity of the dataset brings tremendous challenge for semantic segmentation, for example, the scale variation of objects and long-tail distribution problems. So, we propose a lightweight semantic segmentation framework which consists of a multi-branch feature extraction structure and an adaptive feature fusion module. In addition, we adopt auxiliary supervision strategy to mitigate the side effect of long-tail distribution problem by mining hard samples. Experiments on our dataset show outperforming result compared with existing state-of-the-art lightweight semantic segmentation methods.