Research On Machine-vision-based Fast Crop Detection Method For Robotic Weeder

With the increasing concern on food safety and environmental protection,using of herbicide is receiving more and more limitation.Mechanical weeding has the advantage of less pollution over chemical weeding,while it is more efficient compared with hand weeding.It conforms to the trend of sustainable development of agriculture.Weeding robot is a hopeful solution to achieve high-efficiency autonomous mechanical weeding.Obtaining the positional information of crop plants fast and accurately is the premise of performing autonomous precise weeding operation and one of the key technologies of a weeding robot.Since machine vison systems can output large amount of information in a non-contact way with high accuracy and relative low cost,they are the main approach of in-field plant information acquisition.In this reseach,a machine vision system was designed taking into account the working condition of the weeding robot,planting pattern,and design of the weeding robot.The imaging model of the machine vision system was developed based on pinhole camera model.The result of a test shown that positioning error of the machine vision system was no more than ± 15mm.A software was developed within the Microsoft Visual Studio 2010 programing environment.The software was used as the carrier of crop detection algorithms.It also provided a man-machine interface that supported parameter setting and working information display.As to the crop detection method reseach,three kinds of working conditions were considered when exploring field image processing algorithms for real-time crop detection:(1)According to the working condition of seedling stage weeding for transplanted crops,a fast crop detection method based on pixel histogram was proposed.The main characteristics of crops and weeds at this stage were that crops and weeds had significant difference in size,and crop plants were regularly located while weeds were randomly located.In the proposed method,plants in field images were firstly segmented from the soil background based on their difference in color.Subsequently,crops and weeds were distinguished based on pixel histogram using features of size and location.After that,crop plants were detected and located,and the distances from crop plants to corresponding weeding blades as well as the lateral offset between crop rows and weeding blades were calculated.The crop-blade distances and lateral offset were used for guiding the weeding robot by its control system.Results of field tests under three different lighting conditions indicated that correct identification rates of the pixel-histogram-based method on lettuce,cauliflower and maize were all above 95%.Average processing time for a 640×480 image was 31 ms.(2)According to the working condition in fields with high weed infestation levels that weeds overlapped with each other and differed from crops in color,a fast crop/weed classification method based on color feature and Mahalanobis distance classifier was proposed to achieve efficient crop detection under that working condition.In this method,field images were converted from RGB color space to HSI color space.H(hue)and S(saturation)were used as feature indices.A Mahalanobis distance classifier was build to classify crop and weed pixels in field images.Weed pixels were erased from the field images after the classification so that fast crop detection and positioning could be achieved.Test result showed that by integrating the color feature based crop/weed classification method into the pixel-hisgram-based crop detection method,correct identification rate on broccoli plants in high-weed-pressure field images increased from 80.65%to over 93.6%.Average processing time of the integrated method for a 640×480 image was 112 ms.(3)To solve the problem in practical working conditions that large weeds and volunteer crops occurred near crop plants could interfere crop detection,a fast crop/weed classification method based on convolutional neural network(CNN)was proposed.A CNN with three convolutional layers,three subsampling layers and a single-layer perceptron was built to classify crop plants and weeds in images,aiming at improving the correct identification rate and robustness of the machine vision system to work under complex conditions.Test result showed that the correct classification rate of the CNN on a series of images containing crops and weeds was over 96%.Average processing time of the CNN for a 76×76 image was less than 1 ms.As a conclusion,closely related to autonomous mechanical weeding working conditions,this reseach explored machine vision system design and fast crop detection methods,which laid a foundation for robotic weeding in commercial production systems.