Research On Obstacle Visual Detection Of Tomato Picking Robot

This paper mainly studies the obstacle visual detection technology of Tomato Picking Robot.Aiming at the problem of real-time detection of obstacles by Tomato Picking Robot in greenhouse picking environment,an optimized Mask-R-CNN 101 recognition algorithm was proposed to recognize and segment obstacles Experimental results show that the algorithm has high accuracy and good robustness in the obstacle recognition and detection of Tomato Picking Robot.Aiming at the camera parameter optimization,an adaptive weighted particle swarm optimization algorithm is proposed to optimize the camera parameters,which completes the calibration of binocular camera and improves the calibration accuracy.By using Bouguet stereo correction algorithm and SGBM stereo matching algorithm to complete the stereo matching of the camera,and it meets the detection accuracy requirements of the picking robot.The main research contents of this paper are as follows:Collect pictures of tomato plants in the greenhouse,use data enhancement methods to improve the generalization capability of data sets,the segmentation marking method was introduced to improve the recognition accuracy of tomato vine and stem.Using Labelme tools on the image segmented marking,made into training data set.A new IOU mechanism was introduced to optimize the Mask-R-CNN algorithm and improve the detection ability of vines at the end of stem branches.The tomato vines in the recognition target have many branches and irregular shapes.In order to improve the recognition accuracy,Rest Net101 network is used in the Mask-R-CNN algorithm to form a Mask-R-CNN 101 algorithm with deep network.Using the optimized Mask-R-CNN 101 algorithm to recognize 200 randomly selected pictures in the test set,and the recognition accuracy is 89.65%.There is basically no vine at the end of the stem branch in the recognition result.The recognition effect of Mask-R-CNN 101 algorithm was tested in three environments: forward light,side light and back light.The test results show that the recognition accuracy of segmented markers is 85.32%,83.59% and 80.38% respectively.It is proved that the optimized Mask-R-CNN 101 detection algorithm using segmented markers has high accuracy and good robustness in the obstacle recognition and detection of Tomato Picking Robot.Using Zhang Zhengyou calibration method for Bumblebee2 binocular camera calibration,realize the conversion of the coordinates of the camera coordinate system to world coordinate system.An adaptive particle swarm optimization algorithm was proposed to optimize the camera parameters instead of the LM optimization algorithm in Zhang Zhengyou’s calibration method,and the calibration re projection error is 0.22 pixels,which is half of the error of LM algorithm,it improved the calibration accuracy effectively.The Bouguet stereo correction algorithm in Open CV software was used to correct the images in the left and right cameras,and the SGBM stereo matching algorithm was used to stereo match the binocular vision system.The distance test of the established binocular vision detection system shows that when the positioning distance is less than 1m,the positioning error rate of the binocular vision detection system in the horizontal and vertical directions in space is less than 2.54%,the positioning error rate in the depth direction is less than 0.76%,and the absolute error is within 7.6mm.The error can meet the requirement of obstacle detection accuracy of Tomato Picking Robot.