Intra-row Weed Detection Based On A Binocular Stereo Vision System For The Weeding Robot

Weed control include manual, mechanical, quarantine, biologic, chemical methods, etc.Precision techniques for site-specific weed management has been studied by variousresearchers as it has the potential to minimize the volume of herbicides and to reduceenvironmental pollution caused by excessive use of chemical application. Development of avisual method of detecting intra-row weed under the highly variable conditions remains thegreatest challenge in site-specific weed management.Intra-row weed detection by machine vision poses particular difficulties due to the similarcolor of crop and weed. In this study, an image processing algorithm for crop and intra-rowweed discrimination at V2-V3growth stages utilizing a binocular stereo vision system wasdeveloped and evaluated. The proposed methods use color feature to extract vegetation from thebackground, whilst height and plant spacing information analysis techniques are applied todiscriminate between crop and weeds.Stereo matching is the key step toward height feature acquisition, however, it is generally atime-consuming process. The main objective of this research is to develop a real-time imageprocessing algorithm for intra-row weed detection. The proposed methods use color feature toextract vegetation from the background, whilst height and plant spacing information analysistechniques are applied to discriminate between crop and weeds. During the stereo matchingprocess, correspondence search was performed on edged stereo images and disparity calculationwas applied only to the pixels of edges. This strategy can largely reduce the correspondencesearch range, thereby enhance the weed detection speed and accuracy.The proposed stereo processing algorithm runs at about0.041s, which cuts the runtime by84.4%. The detected distance data inside600mm is accurate within2.5%error. The use ofstereo vision system combined with prior information of plant spacing information provedsuccessful in identifying crop plants with a classification accuracy (CA) of88%, which rendersit suitable to real-time image processing for the weeding robot.