Research On Binocular Vision System Of Sanitation Autonomous Robot

At present,plastic bottles and cans are commonly used as beverage carriers,which are common in streets and lanes.In order to ensure clean roads and good mental outlook of the city,a large number of sanitation workers need to inspect and pick up garbage.Due to the continuous promotion of living and medical conditions,the average lifetime of human beings is increasing,and the cost of employing people in each department is also rising.To address these issue,it is necessary to devise an environmental sanitation autonomous robot binocular vision system that can replace the sanitation workers to identify and locate garbage.In this paper,the core is use binocular stereo vision and deep learning to realize the recognition and location of garbage.The garbage that needs to be identified is mainly divided into two categories: mineral water bottle and pop can.There are two kinds of mineral water bottle and three kinds of pop can.If there are more needs,it is convenient to train other targets to supplement.Firstly,research and analyze the theorem of position,choosing appropriate means,and the calibration test is carried out by using MATLAB.Completing the parallel correction according to the experimental parameters,so that the left and right views can lead the consequence of row alignment.Secondly,the pros and cons of several current matching way are analyzed.According to the experimental results,SGBM is selected as the matching algorithm in this paper.Beforehand,we need to preprocess the left and right photos,which can cut down the effects of noise and the difference in brightness of the two pictures brought,this can promote the exactness of stereo matching.Thirdly,the current mainstream target detection ways is analyzed.According to the demands of precision and instantaneity,the garbage discernment scheme on the strength of yolov4-tiny is selected.Study the principle of YOLO algorithm,on this basis,learn the network structure and loss function of yolov4 tiny,understand its workflow,collect image information for the object to be identified in this paper,make data sets and train the network model,and then use the trained network model to realize the accurate identification of garbage.The central region coordinates of the target garbage in the left view are determined by yolov4 tiny algorithm,and the parallax of the region where the coordinates are situated is fetched to obtain the depth distance and achieve the 3D orientation.At last,the binocular vision system of the sanitation autonomous robot is transplanted to the raspberry pie embedded platform,and the target garbage is placed at different distances to complete the positioning experiment,so as to obtain the best recognition and orientation scope.The actual situation demonstrate that the orientation measure of the environmental sanitation autonomous robot designed in this paper can essentially satisfy the requirements,and can realize the recognition and positioning of the preselected garbage types,with certain accuracy and credibility.The system has completed the identification and orientation of the target garbage in the computer and embedded platform,so it has a certain practical value.