Research On Orchard Navigation System For Tracked Robot Based On Machine Vision And GNSS

The automatic navigation technology of agricultural robots can not only improve the quality of work and production efficiency,but also avoid the harm caused by certain operations and has a significant role in promoting modern agricultural construction in China.In this thesis,an orchard navigation system for the crawler-tracked robot based on combination of machine vision and GNSS was developed.The research project was support by the National International Science and Technology Cooperation and the Guangdong Province Science and Technology project.The main research is as follows:(1)Orchard navigation system hardware and software design.An orchard navigation system based on combination of machine vision and GNSS was developed on a crawler-tracked robot platform Komodo-01.The Qt Creator4.0.2 platform and the object-oriented language C++ were selected to develop the navigation system software under Windows7.Based on the analysis of the characteristics of the crawler robot steering mechanism,a straight-line forward and backward path tracking algorithm based on look-ahead point for a differential-drive robot was developed.(2)Research on visual navigation path recognition algorithm for orchard environment.Based on HSV color space,the fruit tree region segmentation algorithm is proposed and the fruit tree shadow region elimination is achieved by adjusting the model parameters.The traditional mathematical morphology filtering method is difficult to completely eliminate the hole phenomenon caused by the gap of fruit branches and large area weeds on the ground surface in the binary image.To solve this problem,an improved hole filling and noise area elimination algorithm based on scene prior knowledge was proposed.The test results show that the algorithm improves the anti-noise ability of image processing in the orchard environment and effectively detects the fruit tree row area.Due to disturbance factors such as irregular growth of the branches and leaves,using the normal horizontal scanning method in image processing will cause the mislabel tree row edge points.To solve the problem,the prior knowledge of road model was used to make constraints on the normal horizontal scanning results.This algorithm not only improves the accuracy of fruit tree line boundary recognition,but also has certain anti-noise ability.In order to solve the shortcomings of the conventional boundary line detection method,a robust M-estimation boundary line detection method was proposed.The generation of the navigation center path based on identification of both sides fruit tree row and only on side tree row was also studied.Based on the calibration results of the internal and external parameters of the camera,the corresponding relationship between the lateral deviation of image detection and the actual deviation of the crawler robot in real world was derived,which provides a theoretical basis for solving the navigation parameters.(3)Research on end-of-tree row navigation and headland turning for crawler robot.Based on the fact that the tree row areas perceive by the camera was decreased and resulting in visual navigation not reliable when the crawler robot approaches the end of the tree row,an end-of-row detection method based on image processing and end-of-row navigation method based on GNSS was developed.The method can correctly detect the end of row 5m away from the last tree row in the calibration test environment and switch to the GNSS navigation mode.To reduce the headland turning time,space occupied and increase field efficiency,a headland-turning FPID controller for a differential-drive robot was developed and field comparison tests of feedforward,feedback,and FPID controllers were also conducted.The test results show that The FPID balances the capability of the feedforward controller to take preemptive control actions for a heading deviation,while it permits the traditional feedback control loop to provide set-point tracking capability and rejects all other disturbances.(4)To simplified the complex system structure caused by the traditional mediated perception navigation method,an end-to-end navigation system based on monocular vision and deep learning for orchard environment was development.The convolutional neural network consists of five convolutional layers and one fully-connected layer.The network input real time images and output the predictive steering command.To simplify the training sample collection process and reduce degree of human intervention,a neural network training sample collection method based on GNSS crawler robot navigation was proposed.To improve network training and prevent over-fitting,technology such as batch normalization,dropout,image augmentation and ten times ten-fold cross validation was used.(5)Real orchard image sequences were used to evaluate and analyze the image processing algorithm.Straight path and curve path navigation tests of the crawler robot based on the combination of vision and GNSS were tested.The results showed that the robot could successfully go through the artificial tree path with no collisions with the trees and pass through the last row by switching the GNSS based navigation mode.The end-to-end navigation system based on monocular vision and deep learning for orchard environment had also been tested.The results showed that when the camera perceived both sides of the tree rows,the robot moved forward;when the camera could only perceive one side of tree rows,the robot executed the corresponding steering operation;when the camera perceived the-of-row,the robot automatically stopped after driving for 1 meter.