Water Target Recognition Method and Application for Unmanned Surface Vessels

Abstract

Water target recognition is a critical challenge for the perception technology of unmanned surface vessels (USVs). In the application of USV, detection accuracy and the inference time both matter, while it is tough to strike a balance and single-frame water target detection behaves unstable in the video detection. To solve these problems, many strategies are applied to increase YOLOv4's performance, including network pruning, the focal loss function, blank label training, and preprocessing with histogram normalization. The optimized detection method achieves a mean average precision (mAP) of 81.74% and a prediction speed of 26.77 frames per second (FPS), which meets the USV navigation requirements. To build the integrated USV-based system for water target recognition, a water target dataset containing 9936 images is created from offshore USV experiments in which the human-in-the-loop annotation and mosaic data augmentation methods are used. The issues of miss detection and false alarm can be considerably mitigated by cascading the Siamese-RPN tracking network, and the major color of a water target can be retrieved using a local contrast saliency color detection scheme. The system being tested is called 'ME120' includes an embedded edge computing platform (Nvidia Jetson AGX Xavier). Finally, online dataset learning demonstrates the improved YOLOv4 achieves an increase of 66.98% in FPS at the cost of a decrease of 0.79% in mAP when compared with the original YOLOv4 and offline navigation experiments validate that our system achieves high recognition capability while maintaining a high degree of robustness.