6DOF Pose Estimation Of Objects For Robotic Grasping

Vision-based pose estimation of target objects has a wide range of application scenarios,such as the use of robotic arm to complete the grasping of target object,sorting,pose adjustment and so on.At present,there are still many challenges to construct a mature object pose estimation system to assist the robotic arm to grasp.Starting from practical industrial applications,this paper uses sensors to obtain initial color image and depth image data,and utilizes target object model parameters as a priori knowledge to complete the pose estimation of the target object.To solve the problems of weak texture and complex shape of the objects that need to be processed in application,a pose recognition scheme using object contour features and consumer-level RGBD sensors is proposed.When estimating the pose of an object,this paper attempts to change the one-stage process of directly estimating the pose of the target object into a two-stage process.Classifying and determining the basic type firstly,and then obtaining the initial pose state according to the basic type and optimizing the initial pose.The algorithm divides the categories on account of visual surface performance of the object in different poses in advance,and introduces the relationship between categories and the corresponding poses as a priori,fulfilling the effect of obtaining corresponding initial pose utilizing classified category.The regression calculation problem is translated into a classification problem reducing computing difficulty.After successfully obtaining the target object type and corresponding silhouette contour data in application,the model recognizes the category of the current input target object's current state,obtains the corresponding pose data in the prior as the initial pose,and then uses the iterative optimization algorithm to improve estimation results.Classification can be achieved by the most similar template searching strategy based on template matching,or by a classification model based on convolutional neural networks.The most similar template searching strategy based on template matching utilizes customized template data and uses the corresponding template similarity evaluation scheme to maximize the similarity of similar templates while enhancing the ability to filter out heterogeneous template data.The classification model using convolutional neural network extracts features from initial data,and uses the multi-layer perceptron network to achieve classification.The training of the classification model uses the rendered image data.To reduce the difference between the actual image and the rendered image data,the rendered image data is processed to retain the binary contour silhouette data of the target object.In addition,according to the imaging principle,the position of the object during imaging will affect the visible surface of the object in imaging,resulting in an incorrect classification result.For this condition,this paper proposes a compensation scheme for the initial pose obtained by classification using the translation parameters of the object.In order to test the performance of the pose estimation scheme proposed on practical tasks,this paper selects several commonly used pose estimation algorithms with good recognition effects for comparison.When testing the classification scheme based on template matching,three parts used in the industry is used and the number of pose placing the objects is limited,which has a better working effect compared with other methods.When testing a classification scheme based on convolutional neural networks,the initial pose is not limited to several kinds but the target object is placed arbitrarily.In addition to the three commonly used industrial parts,the experimental objects also includes several common objects in the pose recognition database.Compared with other common pose recognition schemes that use texture information,although the classification algorithm discards the texture characteristics of the rendered image,the pose data calculated based on its classification result still has smaller errors and better recognition accuracy rate.