Research On Agricultural Robot Navigation And Weed Identification Based On Multi-source Information Fusion

In this paper,the experimental environment is corn field.Aiming at the problems of easy drift pollution of chemical herbicides and the shortage of agricultural labor force caused by China’s aging population,the navigation line recognition method for corn 3-5leaf stage and high stalk stage and the weed recognition technology between corn rows are proposed and designed a flexible and lightweight weeding robot,which lays a foundation for the autonomous navigation of agricultural robots and unmanned operation in farmland,The specific research contents are as follows:(1)Based on the fuzzy and irregular shape of the path boundary of the corn field,this paper proposes a focused initialization K-Means center robust regression for the 3-5 leaf stage corn field.The navigation line extraction algorithm.First,the super green algorithm and the maximum between-class variance method are used to automatically obtain the green feature binary image,and the morphological processing is combined to improve the image quality.In view of the problem of poor robustness of the traditional method for setting the region of interest,the color threshold range is set in the HSV space transformation,and the mask operation is used to extract the green features to compensate for the information lost by the 2G-R-B algorithm due to interference,and determine the dynamic sense based on the pixel threshold constraint Area of interest.Improve the K-Means initial clustering center selection method,and continuously redistribute the cluster points based on the principle of small initial center point focus,so as to calibrate the center point of the corn plant,use the Thiel-Sen estimation to obtain the navigation line equation,and eliminate the K-Means The interference of abnormal points in the processing results of the algorithm solves the problems of traditional algorithms such as finding the characteristic centroid offset and large calculation amount.Using the information that the height difference between corn and weeds is large,weeds are identified by iterative partitioning and clustering method.Experiments show that the algorithm can adapt to complex environments well and has strong anti-interference ability.Its accuracy rate is as high as 95.11%.It takes 83.6ms on average to process an image with a resolution of 640pixels×480 pixels.The algorithm It provides a reliable and real-time navigation path for the robot to walk in the corn field.(2)Based on the image information of corn rhizomes,a new method for quickly and accurately extracting navigation baselines in the corn field environment at the jointing leaf stage is proposed.First,the image is segmented by 2G-B-R and the maximum between-class variance method,and the image quality is improved by morphological processing,and the vertical projection is obtained by accumulating the denoised image pixels by column.The traditional peak point method needs to set a threshold when searching for feature points,which is time-consuming and has many false feature points.Therefore,a feature point search method based on gradient descent is proposed,which uses a certain point to find the minimum value along the direction of gradient descent.Feature points.According to the principle of corner detection,the false feature points are eliminated by using the gradient changes of the feature point pixels in various directions,which solves the problems of too many abnormal points in the traditional algorithm and the wrong positioning of corn roots and stems.Finally,a random sampling consistent algorithm is used to fit the navigation line.Improve the Canny edge detection operator,increase the calculation of gradient values in two directions,avoid the problem of false edges and missing edges,and finally identify weeds by setting the area of the connected domain.Experimental results show that compared with traditional algorithms,the algorithm can adapt to complex environments well,has strong real-time performance,and is robust even in the absence of seedlings and weeds.The average processing accuracy rate is as high as 92.2%.It takes an average of 215.7 ms to process an image with a resolution of 1280 pixels×720 pixels.This algorithm provides a reliable and real-time navigation path for intelligent agricultural machinery to walk in the corn field.(3)In view of the fuzzy and irregular shape of the path boundary in the corn field,the common field navigation line extraction algorithm will have the problem of excessive deviation in practical application.This paper proposes a camera based on the discrete factor for the 3-5 leaf stage corn field.Navigation line extraction algorithm based on 3D lidar fusion.First,three-dimensional lidar is used to obtain corn plant point cloud data.At the same time,the image obtained by the camera is automatically obtained by using the super-green algorithm and the maximum between-cluster variance method to automatically obtain the green feature binary image,and then the point cloud data after clustering analysis is passed through the lidar and The parameters jointly calibrated by the camera are projected onto the target frame obtained by the image,a multi-sensor data fusion support model is constructed for feature recognition,and finally the acquired feature center point is fitted as the navigation baseline.Experiments show that the algorithm solves the problems of traditional algorithms such as finding feature centroid offset,and has strong anti-interference ability.The algorithm provides a reliable and real-time navigation path for the robot to walk in the corn field.(4)Aiming at the complex environment of corn fields and the situation of farmland in my country,a lightweight,flexible,stable and reliable robot chassis is designed.It has rich interfaces and can integrate a variety of sensors.At the same time,the sensor is calibrated,and experiments and data analysis of the proposed algorithm are carried out.