Position And Attitude Determination Based On Deep Learning For Object Grasping

Nowadays,there are many scenes in industry where object grabbing is done by hand.For example,the placement of sea cucumbers before weighing and grading and before packing are all done by hand.People urgently need a kind of machine system which can grasp and place automatically.In order to realize the automatic grasping system,the robot needs to know the category of the object object,the best grasping position and the grasping angle.In this paper,the problems of grasping position and attitude determination methods at home and abroad are analyzed,such as:the existing methods are not suitable for complex background,low real-time performance and large demand for training data.In this paper,the method of image segmentation based on full convolution neural network is combined with traditional image processing method to determine the position and attitude of object grabbing.In this method,a rotatable bounding box is used to quickly and accurately mark the object detection results for a vision-based robot,and to provide the grasping position and angle for a manipulator mounted on the robot.The specific contents of this paper are as follows:Firstly,object detection methods and data set augmentation methods are studied.This paper analyses and compares the existing object detection methods based on deep learning.Object detection method based on full convolution neural network image segmentation has the advantages of fast running speed and accurate detection results.Full-convolution neural network image segmentation combined with traditional image processing method is selected to achieve object detection.Full convolution neural network has a large demand for data sets and the size of each batch image must be the same during training.The method of expanding the data set is studied,so that the training data set can be expanded online without changing the corresponding relationship between the original image and the label image's pixel value and the spatial position of the pixel.Secondly,the method of object segmentation is studied.Full convolution neural network image segmentation method can overcome the shortcomings of existing image segmentation methods,such as low real-time performance and vulnerability to changes in external scenes.By training and testing the classical FCN-8s network,we find that the accuracy and speed of FCN-8s object segmentation are low and slow.To solve these problems,an improved FCN network model is proposed.The segmentation graph obtained by experiments shows that the accuracy and speed of the improved FCN network model are better than that of FCN-8s.Thirdly,the method of determining the grasping position and attitude is studied.Segmented images have "burrs" and noise.Firstly,the segmented image is processed by etching and filtering.Then the minimum outer rectangle of the outer contour of the maximum connected region on the segmented image is found.The minimum outer rectangle is used to determine the grasping position and angle of the object in the pixel coordinate system.This method has higher positioning accuracy than the traditional boundary frame labeling method,and can realize real-time detection.The disparity map has "holes",which are repaired by multi-level median filling and denoised by median filtering.Using the binocular camera system built in this paper,the grasping position and angle are transformed into the actual space coordinates,so as to prepare the manipulator for positioning and grasping.Finally,the experimental verification of this method is given.A grasping platform consisting of binocular camera system,TX2 and six-degree-of-freedom manipulator is built to verify the effectiveness of all the schemes in this paper.TX2 processes the image captured by binocular camera system and calculates the position and angle of sea cucumber in the pixel coordinate system.The grasping position and angle are transmitted to the manipulator controller through USB serial port to control the manipulator to perform the corresponding grasping action.The experimental results show that the scheme can effectively improve the real-time performance and grasping accuracy.