Research On Vision System Of Cotton Harvesting Robot

Cotton, as an important cash crop, is widely planted in China, especially in Xinjiang district. At present, cotton is harvested mainly by manual and secondarily by machinery in China. However, with the impetus of technology, such as information, computer and automation, agricultural mechanization has completely been achieved in developed countries, and it has been changing to automatization and intelligence, where agriculture is on the way of achieving precise and efficient production. Agricultural robot, a means of achieving agricultural modernization, is being emphastically researched in many countries. Research on mobile harvesting robot mainly includes three parts: design of body structure, autonomous navigation and recognition, location and harvesting of the objects. Vision system, a means of achieving recognition and location of the objects, is especially important. At present, the technology of machine vision has been applied in agricultural automation in many developed countries. However, research on agricultural robot in China is at the beginning, and there is still a big step behind developed countries. Hence, the research of agricultural robot that is a representive of the new agricultural machinery is of significance to hold the opportunity of world scientific and technological revolution in agriculture and promote the development of automation, intelligence of agricultural machinery of China.The objective of this research was to study vision system of cotton harvesting robot. The ripe cotton in the natural outdoors scene was the research object and the recognition and location of cotton fruits were the aims. Simulation and experimental results indicated that harvesting cotton by robot was feasible in theory and practice. Several main points in our research were as follows:1. The color data of cotton fruits, cotton leaves and stems in six color space: RGB, normalized rgb, HIS, YCrCb, I₁I₂I₃ and L^*a^*b^* of the ripe cotton in the natural outdoors scene were analyzed, and RGB color space was considered to be an appropriate space for segmentation. Based on this, the differences of color distribution of cotton fruits, cotton leaves and stems were extendedly analyzed, and (R-B) subtraction module could provide with the best segmentation results. The methods of recognizing cotton fruits were researched, which were based on the combination of maximum Classes Square Error (Ostu) and Freeman chain coding and Back Propagation (BP) neural network respectively. On the combination of Ostu and Freeman chain coding, a new dynamic Freeman chain coding was designed for automatically recognizing the cotton fruits. And the 4-8-1 structure of BP neural network was designed. The accuracy ratio of recognition reached to 86% and 83%, respectively.2. The model of camera was discussed and the results of calibration of the camera were provided. The intrinsic and extrinsic parameters were obtained by the least square method with the pinhole model. The result of calibration showed there were 3.2 pixels errors on horizon and 8.7 pixels in vertical. Based on the parameters of calibration, the method of image rectifying was developed and the distortion of images was rectified by this method. The result showed the method was effective.3. The theory of stereo vision and two models of camera emplacement which were convergent and parallel model were analyzed respectively. The parallel model of stereo vision was selected according to the existing devices of lab. The appropriate measuring distance between cameras was 80mm and the appropriate measuring distance range was from 150mm to 1000mm. Under this condition, the measure error could be controlled within 10mm which was satisfied with the requirements of the manipulator.4. Three matching algorithms based on area, feature and phase often used in the stereo vision were analyzed. And the related definitions and constrained rules used in the algorithms were introduced respectively. With the constraint of the epipole line, with the characteristics of shape gravity and chain coding, the match algorithm of stereo vision was made by the similarity function. Experimental results showed this algorithm was effective and the accuracy ratio of matching reached higher 86%.5. A wavelet field hidden Markov tree structure model based on factor graph was proposed to denoise corrupted cotton images. The model not only used the intra-scale relation of image wavelet field, but also considered the unstationary characteristics of image space field. Simulation results suggested, compared with other image denoising methods, the proposed method had outstanding performance on signal-to-noise ratio and good quality of vision.