Research On Target Recognition And Ranging Methods Of Field Weeding Robots

Field weed damage has always been an important factor affecting crop growth and development.The existing weeding methods usually use manual or spraying of chemical herbicides,which is time-consuming and laborious,and is inefficient.Large-scale spraying of chemical pesticides can also produce residues and endanger the safety of agricultural products.With the development of artificial intelligence and modern agriculture,the use of robots for field weeding operations has become an effective method,which has attracted more and more attention from domestic and foreign scientific researchers.How to distinguish field crops and weeds accurately and efficiently is a prerequisite for robot weeding,and multi-target ranging and weeding path planning have become the key technology.Based on a full investigation of the development of related technologies at home and abroad,this paper has carried out the research on the real-time target recognition and ranging method of the field weeding robot.The main work and conclusions are as follows:（1）A field weed target recognition method based on Faster R-CNN deep network is proposed.Taking images of rapeseed and weeds at the seedling stage under natural environmental conditions as samples,based on the Tensor Flow deep learning framework,a Faster R-CNN deep network model based on different feature extraction networks is constructed;migration is carried out through the deep network model of the COCO data set Training,using Faster R-CNN deep network model and SSD deep network model to share convolutional features,and compare the VGG-16,Res Net-50 and Res Net-101 feature extraction networks respectively.The experimental results show that the Faster R-based VGG-16 The CNN deep network model has obvious advantages in the target recognition of rape and weeds.The target recognition accuracy of rape and weeds can reach 83.90%,the recall rate can reach 78.86%and the₁F value is 81.30%.（2）A multi-target ranging and weeding path planning method for field weeding robots based on depth vision is proposed.Taking images of corn and weeds in the seedling stage of"four leaves and one heart"as samples,the Faster R-CNN deep network model of the VGG-16 feature extraction network is used to achieve real-time target recognition and automatic cutting classification;ultra-green characterization（EXG）parameters for image gray-scale processing,based on the improved OTSU algorithm（IOTSU）to achieve the generation and optimization of the binary image;on the basis of the target contour image of the Canny edge detection operator,a quadratic traversal algorithm（QTA）is proposed to select the two-dimensional coordinates of the target,and the corresponding traversal search box is designed;by mapping the target two-dimensional coordinate point to the three-dimensional coordinate point,the depth camera（Realsense D435i）is used for multi-target ranging and the shortest weeding path planning.The experimental results show that the traversal search box with a size of 100×100 can ensure the accurate selection of the target two-dimensional coordinate points.The search success rate of the secondary traversal algorithm can reach 90.0%on the test data set,which not only effectively saves computing resources,but also avoids A large amount of redundant information generated by the use of depth camera.（3）A real-time target recognition and ranging method based on the integration of target recognition,target ranging and path planning is proposed,and the model deployment and method verification are carried out.Use the depth camera to obtain the key frame images in the video stream in real time,after image preprocessing operations and import the trained deep network model for target recognition and detection,output target classification,target probability and target two-dimensional coordinate points and other information;will be located in the pixel The target two-dimensional coordinate points in the coordinate system are transformed into the target three-dimensional coordinate points in the camera coordinate system,and the distance measurement between crops and crops,crops and weeds,weeds and weeds and shortest weeding path planning are carried out.Finally,through modular programming,the real-time target recognition and ranging model generation API is encapsulated.The model functions are easy to modify and update,and it has strong portability and iterability.