Research on Vision-based Water Surface Garbage Detection and Localization Methods

Abstract

In complex water surface environments, real-time detection and localization of water surface garbage targets are essential for the advancement of water surface cleaning robots. To meet the practical needs of high accuracy and low time consumption, this paper presents a method for identifying and localizing water surface debris using YOLOv8-WSG, which integrates deep learning with vision-based detection and localization techniques. Firstly, relying on the FloW - Img dataset, the water surface garbage homemade dataset was constructed by means of web search, on-site collection, and manual labeling. Secondly, based on YOLOv8, To enhance the detection efficacy for diminutive, overlapping and obscured targets, a P2 sampling layer is incorporated into the neck network, and a small target detection head is introduced into the head network; SIoU is used as the loss function to accelerate the convergence speed and generalization ability of the model. The research results indicate that the method improves 2.2%, 1.7%, and 1.4% in precision, recall, and Map50 values, respectively, under the premise of guaranteeing a detection speed of 1.9ms per frame. And then, the SGBM algorithm is applied to match features between the target regions in the left and right images from the stereo camera, and to compute the disparity and positional coordinates. Finally, the water surface garbage detection and localization system is built on Jetson Xavier NX for validation, and the experimental results indicate that the proposed method achieves an error of less than 0.29m with a detection speed of up to 20 FPS, ensuring high-precision and time-efficient water surface target identification and localization.