Research On Machine Vision Based Object-Locating Technology Of Tomato Harvesting Robot

With the advent of information era, the next wave of agricultural machine will be toward intelligentization. Thus, the researches of agricultural machine will see much broad prospects. This paper makes a comprehensive research on the object locating technology of vision systems of tomatoes harvesting robot, in this research a KH-100 AGV automatic guidance vehicle and a MOTOMAN robot and its control system are adopted, and programs are developed using the Visual C++ 6.0 environment.This paper constructed the binocular vision hardware system of harvesting robot. Firstly, a CCD calibrating approach in the midst of traditional and self-calibrating methods is advanced to separate the inner and the outer parameters. The plane calibration of Mr. Zhang is used to calibrate inner parameter. By extracting feature points in the calibration template, a relatively accurate inner parameter model can be obtained by experiment and calculation. The outer parameter model of two CCD can be constructed by selecting appropriate reference frame and analyzing experiment data. Secondly, the central point based matching method is put forward. The tomatoes' image was analyzed, their surface features were extracted, and each tomato was discriminated using the method of particle segmentation by watersheds; a template based profile emendation method is advanced in which three typical profiles of tomato images are adopted; the coordinates of tomato's central point are obtained; three constraint conditions are advanced and bias ranges are properly selected to match feature points in conjugate images. Thirdly, the paper adopted triangular measurement theory to calculate the 3D coordinates of tomato's central point in conjugate images. Furthermore, based on the comparison between the calculated 3D coordinates and the actually measured 3D coordinates and bias analysis, this paper built a neural net model and the corresponding training system to amend errors. Finally, the system software is written to obtain the 3D coordinates of the rip tomatoes, which can effectively process collected images in natural environment. The experiment indicates that errors can be limited in 0-5mm by adopting the central point match based binocular vision locating technology when the relatively integral profile of a tomato can be photographed by both CCD, thus the acquired accuracy for robots to harvest fruits can be met.