Water-obstacle segmentation based on Swin transformer and sea-sky-line detection

Abstract

Unmanned surface vehicles (USV) has been widely used in the fields of surface quality monitoring and dangerous water exploration, among which the detection and segmentation of obstacles is an important condition for USV to realize autonomous obstacle avoidance. Recently, most of the segmentation methods of water surface obstacles are based on CNN, but due to the limited receptive field of CNN, most of the methods do not achieve satisfactory results in the segmentation of obstacle edge details, sea-sky segmentation and surface reflection. To solve the above problems, we propose a water surface obstacle segmentation network WSSS with sea-sky-line detection as the prior knowledge for feature fusion, Swin Transformer as the dominant backbone and feature pyramid network (FPN) as the feature fusion network. In this network, sea-sky-line detection is used as a preprocessing module to increase the receptive field through the self-attention mechanism and sliding window, and feature extraction is enhanced by FPN, which makes the network segmentation of obstacle edge and water Sky more detailed and reduces the influence caused by reflection and weather. Extensive experiments on the USV datasets MaSTr1325 and Modd2 show that the proposed method WSSS outperforms the current state-of-the-art methods for water surface obstacle segmentation.