The Research On Object Pose Estimation For Robotic Grasping In Cluttered Environment

In a cluttered scene,it is a challenge task for robot grasping target objects in the field of robotics.Especially when the robot grasps the objects in the living scene,it is difficult to recognize and grasp the object with arbitrary pose quickly and accurately.In order to make the robot grasp the objects accurately,it is urgent to improve the ability of object recognition and pose estimation in robot vision.However,traditional object recognition and pose estimation algorithms rely on human-designed feature operators to extract relevant feature information,which are difficult to design accurate feature operators for all scenarios in cluttered environments.And the traditional algorithm cannot estimate the object pose only based on the two-dimensional images.Therefore,the research on object recognition in cluttered scenes and pose estimation based on two-dimensional images is very important to improve the autonomy and environmental adaptability of robots.Different from the traditional methods,the data-driven approach represented by deep convolution neural network(DCNN)has a strong ability to represent the relevant features of the target task through self-learning related feature mapping.Moreover,the equivariant representation of DCNN improves the accuracy of target recognition in cluttered environment.Thus,a series of research on object recognition and pose estimation based on depth convolution neural network is carried out in this paper:Firstly,research on object segmentation based on dilated convolution neural network.In order to solve the problem of small object recognition in cluttered environment,dilated convolution was used to remain abundant spatial structure information in feature maps.An object segmentation network model was constructed based on dilated convolution neural network.labelme was used to generate object segmentation datasets,which is used to train the object segmentation network model.When training the network,the loss function of the model was converged quickly by fine-tuning.In the validation set,the pixel accuracy of the object segmentation network reaches 99.38%.Furthermore,research on object segmentation based on deep residual network.In order to solve the problem of object pose estimation in image,the deep residual network was used to extract more advanced abstract features in the image,and it was established an object pose estimation model based on the deep residual network.Blender,three-dimensional rendering software,was used to render the random pose of object model,so that a large number of synthetic pose datasets could be quickly generated.The position error of the object pose estimation network based on ResNet-18 in the validation set is 1.21 cm and the orientation error is 18.54 degrees.Finally,application of end-to-end real-time object pose estimation system.In order to solve the problem of object pose estimation in real RGB images,a complete end-to-end pose estimation system was built by combining the object segmentation network with the object pose estimation network.At the same time,ROS operating system was adopt to realize real-time object recognition and pose estimation in complex environment.A verification experiment was designed on pose estimation which was based on robot platform.The minimum position error and direction error of the end-to-end real-time object pose estimation system are 1.85 cm and 3.04 degrees respectively.