Research On Visual Navigation Line Algorithm Of Wheat Field Based On Deep Learning

Machine vision navigation is one of the main methods for automatic navigation of wheat field intelligent management equipment.In reality,the environment of wheat field is complicated,and the generation of navigation line is easily affected by environmental background such as light,shadow,weeds,stubble,fallen leaves and wheat seedling morphology.In view of the weak robustness of visual navigation lines extracted by traditional image processing algorithms,this paper studies the extraction algorithm of visual navigation lines for early wheat field operations based on deep learning method,and its main contents are as follows:(1)Study on wheat seedling detection: In view of the tilted morphological distribution of wheat seedling in the image,a R-YOLOv5 wheat seedling detection algorithm based on rotating detection box is proposed.Firstly,on the basis of YOLOv5 algorithm model,angle prediction branch is added.Here,angle prediction is regarded as a classification problem,and the circular smooth label CSL is introduced,which making the detection boxes output rotating according to the distribution of wheat seedlings.And the angle loss is calculated by using dichotomous cross entropy loss function.Then the accuracy of YOLOv5 algorithm model is compared with Faster R-CNN and SSD algorithm models to prove the superiority of YOLOv5 algorithm model.Finally,the improved R-YOLOv5 algorithm model is compared with the YOLOv5 algorithm model for wheat seedling detection accuracy.The results show that the m AP of the improved RYOLOv5 algorithm model is improved by 15.9%,and the output detection boxes could mark the position of wheat seedling row more accurately.(2)Study on wheat seedling row center point extraction: In order to avoid the interference of background factors in the image,an algorithm of wheat seedling row center point extraction based on detection box region is proposed.Firstly,the detection boxes output by R-YOLOv5 algorithm model are clustered self-adaptively,and the detection boxes of the same row of wheat seedlings are divided into a class.Then the gray-scale images of wheat seedlings are obtained by using 2G-R-B,2g-r-b,MEx G,Ex GEx R,NGRDI and S component methods respectively.According to the evaluation parameter e,Ex G-Ex R gray-scale method is selected.Secondly,the noise reduction effects of mean filtering,median filtering and Gaussian filtering are analyzed,and the3×3 Gaussian filtering method is selected according to the structural similarity of evaluation parameters(SSIM).After that,Harris corner and SUSAN corner of wheat seedlings are detected,and SUSAN corner detection method is selected through analysis.Finally,an adaptive center point extraction method of wheat seedling row is proposed.A moving circular window is used to scan in the detection box.During this process,the mean value of corner points in the box is calculated and taken as the center point of wheat seedling row.The extraction results show that this point could accurately characterize wheat seedling row.(3)Study on navigation line extraction: Visual navigation line extraction and position deviation calculation.Firstly,the least square method is used to fit the wheat seedling line.Secondly,in order to effectively evaluate the effect of wheat seedling line extraction,the comprehensive evaluation method of distance error and angle error is adopted.The test results show that the average distance error of wheat seedling line is12.67 pixel,and the average angle error is 1.22°.Then,the visual navigation lines are extracted based on the fitted wheat seedling lines.In order to obtain the accurate position deviation,the camera is calibrated by the Zhang’s calibration method in Open CV.Finally,the position deviation required for automatic alignment of wheat root-cutting fertilization device is calculated by establishing the spatial geometric relationship between coordinate systems,and the average error of position deviation is 14.60 mm.(4)Field test: Field test is carried out on the navigation system developed based on the algorithm in this paper.The experimental platform is tested at different speeds,and the position error of the root cutter path is measured in the field.The results show that the average position errors of the test platform are 11.75 mm,18.56 mm and 31.9mm respectively at the traveling speed of 0.5m/s,1.0m/s and 1.5m/s,which can meet the requirement of not injuring seedlings with root cutting knife.Compared with the algorithm proposed by our research group,the results show that the algorithm proposed in this paper has stronger robustness and higher accuracy,and can meet the requirements of real-time operation.