Research On Ranging And Target Recognition And Positioning Of Spiral Grain Surface Robot Based On Binocular Vision During Grain Leveling Operation

The people depend on food for their livelihood.In the process of grain production,the importance of grain storage cannot be ignored.This link will consume a lot of human labor.At present,there will be a number of uneven grain piles after the harvested grain is stored in the warehouse,which requires manual grain leveling operations,and the carbon dioxide and air particles in the granary will cause harm to the health of the staff in the granary.Therefore,the use of robots to completely replace labor and complete grain leveling operations has become an important issue in the current social grain storage.Since the grain surface of the granary grain leveling robot is an easy-to-flow and easy-to-sink driving surface during operation and driving,the research team designed a spiral walking mechanism as the robot’s walking chassis,and the driving effect is good.In order to realize the fully automatic path planning and operation of the robot in the granary,the first task is to locate the robot in the granary and the grain pile,and accurately find the specific location of the robot in the granary,and the power consumption of the granary robot is large during operation,it is necessary to ensure that the robot maintains the shortest path between the two as much as possible during the process of traveling from the grain surface to the grain pile without obstacles.In view of the above problems,this paper proposes to use binocular cameras as robot sensors to realize the ranging and positioning of robots and grain piles in granaries.The main research content includes:Target detection model: The detection model is summarized and classified,and finally the yolov5 model in the first-stage detection model is selected as the target detection model of the robot and the grain pile to achieve accurate identification on the grain surface.In view of the problem that the grain pile has fewer features and is difficult to identify in the recognition process,the highest weight yolov5 x in the yolov5 model is used for training,and different training rounds are set up for experimental comparison,and the data of the F1 curve and the accuracy-recall curve after training are analyzed,and the detection of different training rounds of models is observed through the test results of the test set.Finally,the average accuracy mean m AP under different rounds was obtained.Through multiple attempts,it was found that when the training round is 300,the average accuracy mean m AP can reach 99.5%,which meets the accuracy requirements of yolov5.The recognition effect of the model can be tested on the spot through the camera.Binocular vision system: Select the binocular camera as the camera equipment used in the experiment,according to the imaging principle of the pinhole model,select the Zhang Zhengyou calibration method to calibrate the camera,and consider the distortion problem of the camera,and solve the camera’s internal parameter and external parameter matrix.The binocular camera is corrected by the polar line correction method,so that the corresponding points of each pixel of the left and right cameras are matched.According to the principle of parallax of binocular cameras,compare the advantages and disadvantages of the BM algorithm and the SGBM algorithm,and choose the SGBM algorithm as the parallax algorithm of the robot camera.The image preprocessing is carried out through the WLS least squares algorithm and the median filtering algorithm to reduce the noise interference of the image.The depth information of the object point is obtained according to the ranging principle of the binocular camera,that is,the ranging model of the camera is obtained through the depth information of the robot and the grain pile.At the same time,the spatial position information of the target object can be obtained according to the parallax algorithm.Model fusion ranging: By improving the binocular camera model,Euclidean distance is used to calculate the relative position information of the robot and the grain pile.In the pycharm software,the PIL image processing environment library is used to integrate the binocular camera model with the yolov5 model to explore the applicability of the two models.The experimental results found that the two models can be fused,and the applicability of the model is high.The grain surface information of the test bench is obtained by setting the height and placement angle of the camera,and the relative position measurement experiment of the robot and the grain pile is carried out.Through the results of the ranging experiment,it is found that the deviation between the ranging result and the actual distance is below the threshold value of 10 cm.Hardware and software design: stm32F407VET6 was selected as the microcontroller of the robot in terms of robot hardware.The Jetson Xavier Nano produced by Nvidia was selected as the Ubuntu system computer.The camera is selected as the Mesco MSK-SM model binocular camera.On the software side,the Ubuntu system computer is used to deploy the fused model into the robot’s ROS system.In order to communicate between the upper computer of the robot and the lower computer in future research.Through the QT library of the python language,the binocular camera fusion model of the robot is encapsulated into a visual GUI interface,and the GUI interface is deployed to run under the QT interface of ROS.It can be convenient for future researchers to use.Robot visual positioning experiment: Debug the equipment before the experiment,put the robot into the test bench,run the fused model,and randomly set four grain piles on the grain surface,set the origin position to be in the geometric center of the grain surface,and obtain the three-dimensional static position of the spiral robot on the test bench and the grain pile through the binocular camera.By correcting the camera’s image sensor coordinate system,the origin is in the lower left corner of the test bench,and the edge of the image fits with the edge of the test bench for two-dimensional positioning.Five points are selected on the grain surface.In addition to the original position of the robot and the position of the four grain piles,the remaining three points are randomly selected positions,and the three-dimensional dynamic positioning of the grain surface is carried out.For the problem of low two-dimensional dynamic positioning accuracy,the UKF traceless Kalman filter is used to optimize,and the average deviation and root mean square error RMSE before and after the filter are calculated.The results show that during static positioning,the RMSE of the three-dimensional positioning coordinates and the RMSE of the two-dimensional positioning are less than the threshold value of 10 cm,and the accuracy rates are 92.5% and 93.6%,respectively.During dynamic positioning,the RMSE of the three-dimensional positioning coordinates and the two-dimensional positioning coordinates filtered by UKF Kalman are both less than the threshold value of 10 cm.The accuracy rates were 91.2% and 92.2%,respectively,which met the positioning requirements of the grain surface robot.