Target Detection Algorithm of Forward-Looking Sonar Based on Swin Transformer

In recent years, with the deepening of the exploitation of marine resources, the demand for marine environment exploration is also increasing. However, due to the complex and changeable marine environment, the exploration risk of underwater manned spacecraft is relatively high. Autonomous underwater vehicle (AUV) has become an essential tool to explore and exploit marine resources. During the execution of underwater tasks, AUVs often encounter collision and entanglement of various obstacles, which will pose a fatal threat to AUVs. Forward-looking sonar (FLS) is one of the main sensors used by AUVs to detect underwater targets. By continuously transmitting sound waves and receiving echo signals, AUVs equipped with multi-beam forward-looking sonar can generate real-time visual field acoustic images, and then detect the images with target detection algorithms to solve the above problems. However, the visual acoustic image has the characteristics of more texture feature information and less semantic feature information, and it requires real-time reasoning. Some existing classical target detection algorithms are mostly designed for some data scenes with strong semantic meaning, and cannot be well adapted to the visual acoustic image. Therefore, a new target detection algorithm, Swin_FLS, is proposed in this chapter. The improved Swin_FLS can more fully extract the texture features in the forward-looking sonar image and adapt to the characteristics of the forward-looking sonar image data set. Finally, 85.2 mAP and 14.1 FPS are obtained in the test set. It surpasses the accuracy of some classical algorithms directly applied to the forward-looking sonar image data set.