Research On Target Detection And Localization Technology Of Picking Robot Based On Binocular Vision And Deep Learning

The National Development Program "Made in China 2025" requires for improving information collection,intelligent decision making and precision operations in the agricultural production.Intelligent harvesting of agricultural machinery and equipment will promote the further development of China's agriculture to intelligent agriculture.In the face of complex natural environments,the existing technology can not enable the picking robot to detect accurately and locate the fruit in real time.In view of the above problems,the paper takes kiwi as the research object and studies the target detection and positioning technology of picking robot.The main contributions are as follows:First of all,due to the cumbersome steps of camera calibration,it is not conducive to the rapid batch operation of the engineering team,so based on the relevant theory and Opencv to obtain the ViEye camera calibration tool.The internal parameters,distortion parameters and external parameters of the binocular camera are obtained by calibration.The image obtained by the left and right camera is corrected to the same plane,and the calibration error is less than 0.05%.Secondly,in view of the low recognition accuracy and time-consuming of kiwifruit detection methods,a kiwifruit detection model based on Retina-Net is established to detect kiwifruit in the natural environment:Under the framework of Pytorch,ResNet-50 was used as the feature extraction network by making data sets,selecting appropriate parameter-adjusting strategies and optimization functions.The average recognition precision is 91.35%,and the recognition time of a single image is 0.08s.The detection accuracy and speed are greatly improved,which can provide data support for the three-dimensional positioning of fruits.Then,in view of the difficulty of extracting the feature points of kiwifruit in the natural environment,the YUV histogram equalization and SURF algorithm are used to enhance and extract the features of the image,and epipolar constraint is used to eliminate mismatching points.Based on the above research,A stereo matching algorithm for kiwifruit in the natural environment is proposed to complete the stereo matching of the kiwi fruit picking area,and the matching accuracy is 89.39%.Finally,a three-dimensional positioning experiment is carried out on the kiwifruit picking center point,and the positioning error is less than 3.5%.It meets the requirement of three-dimensional positioning for the kiwi picking robot.In summary,the paper takes the kiwifruit in the natural environment as the research object,and studies the target detection and positioning technology of the picking robot based on binocular stereo vision and deep learning,and realizes the effective detection and three-dimensional spatial positioning of the kiwifruit.