Research On The Method Of Recognition And Location Of Scattered Workpiece Based On Multi-data Fusion

Robotic grasping has always been the focus of research around the world.The variety of object types,the arbitrary placement of object poses,and the insufficient of robotic perception and judgment are the bottleneck problems on the robotic grasping and object classification.At present,deep learning technology has been successfully applied in different fields,and the use of large-scale data has greatly improved the performance of the technology.This paper faces the recognition and positioning of scattered workpieces,using deep learning technology,fusing multi-mode sensing datas,and establishing the recognition classifier and grasping posture detector respectively.Main work:The research on establishing the recognition classifier of workpieces.In this paper,The public data set is preprocessed and extended to obtain the data sets to be classified.Due to the small number of data sets,this paper uses the method of transfer learning to train the neural networks VGG16 and Inception-v3;By analyzing and comparing the test results of the two classification models on the test sets,the Inception-v3 neural network model is finally selected as the artifact classification model.The research on establishing the grasping posture detector of workpieces.When the object grasping rectangles are obtained only using the RGB images,the accuracy of the object grasping rectangles are easily affected by the color and background of the objects.Therefore,for RGB-D data,this paper constructs a multi-mode data fusion mode to process RGB-D information.First,two single modality CNN models are built on color images and depth images respectively.Two grasp detectors are used as feature extractors to extract the color and depth grasping features.At the end of the model,two grasping features are fused,and an output classification layer is added to realize the construction of multi-mode grasping rectangular detector.Experimental verification of online classification and grasping detection.The Kinect v2 camera is used as the main sensor to obtain the scene data of the target workpiece,complete the calibration of camera and robot.The random sampling consistent algorithm is used to realize the segmentation of the target workpiece and detect the optimal grasping rectangle.The top-1 grasping rectangle is not necessarily the optimal one,so this paper adds a center-of-gravity comparison algorithm to detect the optimal grasping rectangle.Formulate corresponding rules to map to the actual crawl parameters.Set up the experimental platform,place four different kinds of workpiece on the table with five different attitudes,carry out the research of classification and grasping posture detection.The experimental results show that this system is effective.