Research On Visual Perception Method For The Working Condition In Paddy Field

The automatic driving operation of agricultural machinery in paddy field is an important research content of intelligent agriculture,among which the perception of navigation information and operating condition is the key point to research.At present,the proposed extraction methods of visual navigation line are mostly based on traditional image processing algorithms,which have poor robustness in complex environments.In recent years,the application of deep learning algorithm in the field of machine vision has gradually matured,and it is suitable for feature extraction in unstructured environment.At present,the researches on the automatic operation of agricultural machinery in outdoor field mainly focus on the navigation,however,there are few researches on the perception of the working condition of farming such as rotary tillage.Aiming at the complex unstructured environment of paddy field,this paper adopts machine vision as the core method to carry out research on visual navigation line extraction and detection of rotary tillage operation condition in paddy field.The main research contents are as follows:(1)Aiming at the disturbance caused by the complex lighting environment,weeds and plant absence in paddy field,a visual navigation line extraction algorithm based on Fast-SCNN semantic segmentation network was proposed to extract the center lines of rice seeding rows.Pictures of paddy field at different periods after transplanting and under different lighting were collected,and pictures at 1 week and 3 weeks after transplanting were selected as the research objects and data sets were made based on the pictures.To separate the seeding rows from the image,the study trained the semantic segmentation network to select the optimal model.In this study,Fast-SCNN lightweight semantic segmentation network was selected by comparing the effects of four different semantic segmentation algorithms.By the segmentation of Fast-SCNN network,each row of seedlings in the seedling image at different growth stages formed a complete connected region.The MIoU of segmentation result is 82.32%,the pixel accuracy is 90.16%and the average time spent per frame is 24.04ms.(2)In the resulting images of semantic segmentation,an outer contour extraction algorithm based on boundary tracking was used to extract the contours formed by the connected regions of each seedling rows.By comparing the graying results of four graying operators,2G-R-B operator was selected to gray the original color image,then OTSU algorithm and median filter were used for binaryzation and denoising.Corner points in the binary picture were extracted as feature points of seedlings using the improved FAST corner detection algorithm.Feature points were classified based on the extracted contour information,and each category of feature points was fitted into a straight line by Hough transform algorithm based on known points.The experimental results show that the average angle error of rice seedling row centerline extraction is 1.46°,the average time spent of each frame including semantic segmentation is 158ms,and the accuracy rate is 95.9%,which meets the real-time and accuracy requirements of visual navigation in paddy field.(3)A method for detecting rotary tillage conditions based on machine vision was proposed for rotary tillage operation in paddy field.This study selected 2 working conditions,residual stubble amount and uniformity of tillage depth as research objects.The residual stubble amount detection method based on the standard deviation of Y component of YCrCb color space and the uniformity of tillage depth detection algorithm based on the depth information of rotary tiller were proposed,which could detect the working conditions of rotary tillage in real time.By field tests,the judgment results of the algorithm in this paper were compared with the results of artificial judgment,to verify the accuracy of the methods proposed in this paper.The accuracy rates of the proposed methods in the two conditions are 83.6%and 81.0%.This study provides a real-time and effective method for detecting rotary tillage conditions in paddy field.The research realizes the transformation from qualitative evaluation to quantitative evaluation of tillage quality,which lays a foundation for the digitization of tillage quality.Results of experiment show that the proposed algorithm based on Fast-SCNN semantic segmentation network and improved FAST corner points can accurately extract seedling row center lines in real time,with achieves an accuracy of 95.9%.The working condition perception algorithm of paddy field provide in this study can perceive the working condition of rotary tillage in real time.