An Efficient Network for Obstacle Detection in Rail Transit Based on Multi-Task Learning

Obstacle detection, as one of the core functions of autonomous trains, consists of two specific tasks: track area segmentation and foreign object detection. Conventionally, the two tasks are performed separately. In contrast, we first propose a unified and efficient multi-task network to handle these tasks jointly. Specifically, a schema with one encoder for feature sharing and two decoders for specific tasks is adopted, and three designs such as effective Squeeze-Excitation module are proposed to obtain well shared representations. Furthermore, we establish a high-quality dataset by fusing RailSem19 and MRSI. Experiments demonstrate that our method obtains competitive performances on both track area segmentation and foreign object detection compared with state-of-the-art methods while far exceeding the existing approaches in terms of computational efficiency.