Research On Field Maize Seedling Recognition Method Of Weeding Robot

Weeds in the field are one of the main hazards of maize seedling growth.Weeds usually have strong vitality,grow and reproduce quickly,competing with field crops for soil nutrients,water,and sunlight,affecting the growth and development of maize seedlings and leading to reduced yield and quality of maize.With the continuous development of artificial intelligence technology,the application of machine vision technology in agricultural intelligent equipment is becoming more and more widespread.Using machine vision technology to identify maize seedlings can significantly improve the operation of weeding robot equipment in maize fields.However,due to the complex field environment,the current weeding robot fieldwork generally has problems such as slow detection and imprecise identification of maize seedlings,which can easily cause maize seedlings injuries and wrong removal during work.For this reason,this thesis proposes a weeding robot field maize seedling recognition method,and the main work and conclusions are as follows:(1)A field maize seedling and weed detection method is proposed.The images of maize seedlings and weeds in a natural environment were collected as samples,and the mainstream target detection models as well as various versions of YOLOv5 were compared,and the target detection model based on YOLOv5 s was selected and trained to detect maize seedlings and weeds in the field.The accuracy rate,recall,and m AP@0.5 of the model during training respectively reached 87.6%,89.4%,and 90.5%.From the figure of experimental detection results,we can find that the trained YOLOv5 s model can effectively detect maize seedlings and weeds in the field.(2)A field maize plants contour extraction method is proposed.In order to accurately locate the contour shape of maize plants,the HSV color space model and RGB channel separation were used to extract the green leaf and root features of maize plants,respectively.After obtaining the green leaf and root images,the F-B algorithm was used to select feature points and describe and match the feature points,and the random sampling consistency algorithm was used to eliminate the incorrect matching points and stitch the complete maize plant images,and then the Sobel operator was used to extract the plant image contours.The experimental results show that the F-B algorithm has improved matching speed and accuracy compared to scale-invariant feature transform(SIFT),speeded up robust features(SURF),oriented FAST and rotated BRIEF(ORB)algorithms,with a matching accuracy of over 80%.Using the Sobel operator to extract plant image contours,the obtained contours have better clarity and completeness and can achieve the extraction of maize plant contours with faster speed and higher accuracy.(3)A field maize seedling identification system is designed.The system is deployed on a weeding robot and uses a depth camera to acquire key frame images from the video in real-time as input,which is imported into a trained maize-weed detection network model for detection.Then,the target maize seedling is segmented to obtain the target maize seedling image.Further,the target maize seedling images are contour extracted.Practical tests have demonstrated that the system achieves rapid and effective identification of maize seedlings in the field.