Research On Vision System Of Cucumber Harvesting Robot

China is a large agricultural country; agricultural technology and equipment are related to the national economy’s competitiveness and sustainable development of the agricultural economy. To further enhance agricultural mechanization and automation levels and to reduce production costs, more and more attention has been paid on agricultural harvesting robots research world widely and it has been become an important research direction of agricultural machinery and equipment.Cucumber is a kind of world vegetables with a long cultivation history. In China, cucumber harvesting and classification are mostly done by labor, which are inefficiency and high labor intensity; human factors can bring different criteria for classification, which is difficult to meet the needs of modern agriculture. Agricultural mechanization has entered the fast lane; advanced equipment has been used in all aspects of agricultural production and achieved intelligence automation. This is consistent to requirements of precise and efficient modern agriculture. Therefore, it is necessary to explore and research cucumber harvest robot. The one hand, it can improve labor’s operating environment, low labor intensity; the other hand, it can increase picking efficiency and protect the natural environment.The research focus on agricultural engineering-machine vision technology, the key technology of cucumber harvesting robot vision system-object recognition and navigation path extraction were systematic studied, the main contents include following aspects:(1)Spectral characteristics of cucumber and their stems、leaves in the near infrared spectral were studied. From the experiment significant differences of the contrasting reflectance spectra of cucumber and stems、leaves was found. principal component scoring and Mahalanobis distance calculation were used to eradicate abnormal samples, cross-validation was made for sample modeling, which resulted into7principal components and a PLS model was established. The prediction of the pre-set validation sample set confirmed that the recognition rate up to100%. (2)Color space conversion and grey scale image enhancement processing was performed on greenhouse cucumber, based on its image characteristics. Statistical analysis of Chromatic aberration on the three kinds of cucumber image color space (9color channel) shows that RGB color space is suitable and to be the basis of image processing. In image enhancement process, pulse coupled neural network time matrix was employed to enhance cucumber gray scale image. This method combines the visual characteristics of human eye and Weber’s law, which effectively enhance cucumber image. The results showed an enhanced image highlighting the image contrast and preserve the image detail.(3)Segmentation method of greenhouse cucumber image was investigated. For cucumbers with similar color to the stems and leaves, improved pulse coupled neural networks was employed to segment cucumber image, which coupled the gray information and spatial information of cucumber image in connection coefficient, and adaptive adjustment of the parameters was made to segment cucumber image with dynamic threshold. When image’s two-dimensional Tsallis entropy achieves its maximum, the maximum Shannon entropy as well as the minimum Cross entropy was adopted in the processing, which is named as the iteration of PCNN in this paper. The statistical results also shows that the two-dimensional Tsallis entropy terminating criterion of PCNN can obtain good segmentation results.(4)Cucumber target identification strategy in greenhouse environment was investigated. Based on the binary image of pulse coupled neural network segmentation, mathematical morphology was employed to further process the binary image. Four geometrical features and3texture feature values based on gray level comatrix were extracted from each connected region in binary image, these values were used as input feature vectors to train LS-SVM classifier for distinguishing the cucumber target in image. The experiment results showed that correct identification rate up to82.5%in70cucumber images, indicating that pulse couples neural networks combines with LS-SVM method was suitable for recognizing cucumber in complex background greenhouse.(5)A generalized fuzzy Hough transform process was proposed for the partially shielded cucumbers. Bezier curve was applied to fit the central axis of cucumber. The idea of generalized column was employed to describe the cucumber. Double points of binary cucumber image were extracted, whose invisible angle was used as an index, the ratio of distance of double points and the distance of double points in template to determine the scale of cucumber image. Also fuzzy concept was applied to the voting mechanism to reduce false voting. Generalized Hough transform was used for calculating the potential location of cucumber target. By comparing the "area of difference" between the cucumber and the different rotation angle template the position and orientation of target and the restoration of the cucumber shape was achieved. Results confirmed that the generalized fuzzy Hough transform is best to identify the part of block and the different positions and orientation of cucumber target.(6)The navigation path extraction algorithm for cucumber robot in unstructured greenhouse environments was investigated. Column scanning of cucumber image captured by the CCD accumulated gray value of each column was performed to determine the possible regions of navigation path. Then the color differences of the plant and the road in image was analyzed. The three components of RGB image was were computed (EXG and EXR) to acquire gray image with bimodal characteristics histogram. OTSU was used for image segmentation and progressive scanning was carried out from the possible navigation path in both left and right direction on the images. Discrete point of navigation was obtained for getting the average of horizontal and vertical distances from the mutual points on each gray scale. Values of discrete point obtained by taking the average of the continuous five discrete points and final least square method used to fit the navigation straight line.