Improved Object Detection Method for Unmanned Surface Vehicle Using Real-Time Neural Networks

Real-time and accurate object detection is a critical (b) prerequisite for Unmanned Surface Vehicles(USVs) to perform intelligent tasks based on images or videos. While the maritime environment always encountered various extreme scenarios, such as rainy or foggy weather, strong lights and far vision, which all seriously harmed the performance of state-of-the-art methods for normal object detection when directly applied them on USV. Therefore, we proposed an improved object detection method for USV based on Yolov4, which focus on repairing the performance loss caused by the unfavorable factors under maritime environment. Firstly, we adjust the default anchor size in ordinary model which helps detect the tiny object from a far vision, as well as the anchor ratio, fitting the shape of ships more to implicitly improve the detection precision. Secondly, we take full advantage of data augmentation to increase the robustness of object detection under extreme brightness. Finally, we enriched our training data with more rainy and foggy images token from different maritime scenes which enhanced model’s ability to detect objects under extreme weather. Extensive experiments demonstrates that proposed improved method effectively achieved real-time and accurate object detection for USV under maritime environment.