STOS-Net: Spatio-Temporal Contextual Obstacle Segmentation Network for Surface Vehicles

Robust navigable segmentation and obstacle detection are crucial for an autonomous surface vehicle. In this paper, to effectively tackle the severe challenge of distinguishing obstacles from surface disturbances by virtue of inter-frame correlation, a bilateral spatio-temporal contextual obstacle seg-mentation network (STOS-Net) is innovatively devised. In the context branch, to capture spatio-temporal contextual information, by utilizing the shared encoder to extract features involving previous and current frames, a spatio-temporal attention module is proposed, such that potential relationships among features of consecutive frames can be effectively established, and thereby distinguishing surface disturbance and actual obstacle. Subse-quently, the Sobel edge detection operator is deployed to generate boundary maps, which guide the detail branch to produce more precise edge feature maps, thereby significantly enhancing the accuracy of waterline segmentation. Moreover, to address the significant multi-scale variations caused by the size of obstacles, the multi-scale information from the two branches is integrated by a multi-layer perceptron module. The findings demonstrate that the innovative STOS-Net attains an F1 score of 93.5 on the MODS dataset, along with a localization error of 12.1 pixels for waterline segmentation.