Autonomous Water Surface Garbage Detection Based On Computer Vision

These days with the improvement of urban areas all through the world waste management is becoming a significant problem.Most of the methods that are currently applicable find it hard to deal with the volume of solid waste produced by the developing metropolitan populace.The current technique for following waste containers in different regions is inflexible,wasteful,and hard.Observing and monitoring the evaluation of garbage areas in metropolitan scenes depends on manual investigation and photographic record,which makes it a troublesome and tedious undertaking.During the review cycle,human mediation and bulkiness issues frequently occur..Different from other kinds of objects like a vehicle,pedestrian,trash has no moderately clear definition.Because the judgment of trash consistently has certain subjectivity,in various circumstances,it will deliver diverse judgment results.Since the a decent variety of scenes where trash shows up,the accuracy of test results will be influenced.Nowadays with the development of big cities,it’s important to provide an automatic garbage detection method to help urban garbage problems.This paper focuses on the research of the detection algorithm of the surface garbage and applies computer vision technology to the detection of the surface garbage to realize the automatic detection of the surface garbage.This paper introduces the latest development of research on surface water garbage target detection.Aiming at the non-water surface area and water surface interference factors in the images obtained by the cleaning robot in operation,the surface garbage detection is divided into three stages: water surface segmentation,garbage target detection,and segmentation detection fusion.The first stage of water surface segmentation,using a mobile net v3-large(1.0)based.The water surface segmentation algorithm is composed of a lightweight neural network and an improved Atrous spatial pyramid pooling(ASPP)semantic feature fusion module.The image obtained by the cleaning robot is divided into water and non-water areas.In the second stage,garbage target detection is carried out in the standard version of you only look Once＿V3(yolov3)target detection algorithm,and to improve the speed and accuracy I used 1)generation intersection over Union(GIOU)2)improved feature fusion layer and 3)network slimming technique.The algorithm prunes the network to improve the detection speed of the network;in the third stage,the water surface segmented from the image in the first stage is fused with the garbage target detected in the second stage,to eliminate the garbage in the non-water area and realize the water surface garbage target detection.Through experimental verification and testing,the average pixel accuracy of the proposed water surface segmentation network is 95.9%,and the average intersection over union is 94%.At the same time,the segmentation speed of 56 fps is achieved.Compared with the standard version of yolov3,the average detection accuracy of the improved yolov3 garbage target detection network is improved 4.8%,the detection speed is improved nearly twice.The detection accuracy of the whole water surface garbage detection algorithm after segmentation detection fusion can reach 77% on the data set with only water surface garbage target and has the detection speed of 32 fps.The experimental and test results show that the proposed method can detect the garbage target on the water surface in a complex environment,and has a certain engineering application value.