Comparison and Simulation of Deep Learning Detection Algorithms for Floating Objects on the Water Surface

Abstract

Cleaning up floating waste on the river surface is an important means of preventing ecological pollution and slowing down the growth of marine waste. Unmanned boats combined with real-time object detection algorithms can provide solutions for cleaning floating garbage from rivers. However, the detection accuracy and response speed of different object detection algorithms vary greatly when dealing with different complex scenarios. The study uses five classical object detection algorithms: Faster RCNN, Cascade RCNN, Yolo v3, Yolo v5-s, Yolo v5-m, and FloW-Img, a river floating trash dataset, to experimentally verify and compare the detection accuracy and efficiency of each algorithm. The experimental results show that, in terms of detection accuracy, the Yolo v3 model has higher detection accuracy than the remaining four algorithm models, with mAP@[IoU=0.5] of 0.924 and mAP@[IoU=0.5:0.95] of 0.536; and in terms of model parameter scale, the Yolo v5-s model has the smallest parameter scale of 13.36Mbit; in terms of detection rate, the Yolo v5-s model has the fastest detection rate, with FPS of 78.13 in the server environment and 31.3 in the embedded simulation environment. The results of the above experimental indexes show that the Yolo v3 algorithm model is suitable for inland floating object detection scenarios with high accuracy requirements; Yolo v5-s and Yolo v5-m have smaller parameter scales and are more suitable for embedded device detection applications.