AUV target perception in forward-looking sonar images: A low-pixel dataset and multi-frame segmentation network

This study investigates low-pixel segmentation of a moving autonomous underwater vehicle (AUV) observed by a fixed forward-looking sonar (FLS) and formulates the task as a pixel-level segmentation problem under low-pixel and noise-interference conditions. We release the Underwater Sonar Moving Target (USMT) dataset, which contains 21 temporally ordered sequences and 7,136 annotated FLS frames collected using a fixed FLS while observing a single AUV in a controlled lake environment. To exploit motion cues across consecutive frames, we propose LSNet, a multi-frame segmentation network adapted to the characteristics of sonar imagery. LSNet incorporates three improvements: (1) a spatio-temporal scale fusion block that combines multi-scale spatial features with variance-based temporal motion cues; (2) a decoupled 3D spatio-temporal attention module that separates local perception from temporal verification; and (3) a Muon-Adam hybrid optimizer with a joint loss function. Technical validation shows that the rectangular B-scan representation is more suitable for segmentation than the sector-shaped representation. On the USMT dataset, LSNet achieves an IoU of 68.21%, outperforming the retrained LVNet reference by 19.57 percentage points and the strongest additional comparator, LMAFormer, by 5.45 percentage points, while maintaining an inference speed of 91.23 FPS. These results demonstrate the effectiveness of multi-frame modeling for low-pixel AUV target segmentation in external FLS-based monitoring.