| The vision navigation technology of cropland is an important part of intelligent agricultural machinery.However,there are factors such as duckweed,cyanobacteria,water surface reflection,strong wind,variable illumination conditions,rice seedlings at different growth stage and so on that have effect on the accuracy and efficiency of autonomous navigation of paddy field weeding robot.Therefore,to solve those problems,a vision navigation method based on YOLOv3 for paddy field weeding robot is proposed.The main research contents are as follows:(1)Research on rice seedlings detection: A method based on the depth neural network model of yolov3 is proposed for detecting the rice seedlings in complex environment paddy filed.Before making training set of images,the region of interest(ROI)of the rice seedlings images was constructed to extract the rice seedlings with guiding significance for navigation and which will help reducing the datas of subsequent image processing.In the process of making training set,this paper proposed a way to labeling the rice seedlings row segmentally to improve the accuracy of the ground truth boxes.In the training process,to improve the performance of the model,the image data enhancement is carried out and the priors anchors size are calculated by K-means algorithm.After the training,the best model and the optimal confidence threshold were selected according to the evaluation index of model.The experiments results show that the average accuracy of the detection algorithm is 93%,and the average detection speed of each image is 40.56 ms,which proves that the detection algorithm can meet the real-time and accuracy requirements of navigation.(2)Research on rice seedling feature extraction: For the purpose of reducing the influence of background noise on the extraction of navigation lines,a new method of rice seedling feature extraction within the bounding boxes of the same rice seedlings row is proposed.Firstly,the output bounding boxes after rice seedlings detection is clustered adaptively.Secondly,in the process of rice seedlings feature extraction,5 kinds of commonly used gray scale methods including 2G-R-B,Ex G-EXR,MEx G,NGRDI and S component are compared and analyzed.Thirdly,3 kinds of corner algorithms including Harris corner,FAST corner and SUSAN corner are used to extract the feature point within the bounding boxes of the same rice seedlings row and compared in terms of extraction effect and calculation time.The results show that SUSAN corner is more suitable for feature point extraction of rice seedlings.Finally,Hough transform based on the known point and least square method are used to extract the navigation lines.And after that,the navigation parameters are calculated by the spatial geometric relationship between the established coordinate systems.(3)Experimental studies: The vision navigation system is tested.The positioning error of the vision navigation algorithm is measured.Through the navigation experiment of the weeding robot at different speeds,the average and maximum values of the position deviation and angle deviation of weeding robot during navigating are obtained.Besides,the performance of the navigation algorithm of this paper is compared with that developed by our group.The comparison results show that the navigation algorithm of this paper has higher accuracy,better adaptability and robustness to the complex paddy field environment of South China.The research and experiments show that the algorithm proposed in this paper has better robustness,faster detection speed and higher accuracy for the complex paddy field environment compared with the existing algorithm.The algorithm proposed in this paper can meet the real-time and accuracy requirements of robot vision navigation. |
PDF Full Text Request