Research On Object Six-degree-of-freedom Pose Estimation Method Based On Improved Fully Convolutional Neural Network

Six-degree-of-freedom pose(6Dof pose)information is a kind of spatial information,which is an important input in the fields of augmented reality,robotic arm grasping,and automatic driving,and has received widespread attention.The existing 6Dof pose estimation technology is highly dependent on additional sensors,and it is difficult to restore the attitude of the target object stably and accurately under conditions such as obvious light changes and severe occlusion.In order to solve the above problems,this thesis designs a pose estimation system,which takes a single color image as input and the six-degree-of-freedom attitude of the target object as output,aiming to provide high-precision attitude information for various downstream technologies.The main research contents are as follows:(1)Aiming at the problem that it is difficult to estimate the accurate pose of the object due to the information dimensionality reduction generated during the mapping of 3D objects to 2D images,this paper designs a two-stage pose estimation method,extracts 2D-3D feature point matching pairs through a fully convolutional neural network,and restores the feature point matching relationship to the 6Dof pose of the object based on EPn P solution.(2)A multi-source fusion feature point is proposed to fuse the surface feature points and 3D bounding frame feature points obtained by FPS sampling,which reduces the sensitivity of the system to texture information and improves the stability in the pose estimation task for different objects.Through the principle of small hole imaging,the selected 3D feature points are mapped to the 2D image,and Gaussian distribution is applied to obtain a heat map,so as to retain the spatial information as much as possible and complete the data marking.(3)Aiming at the feature point extraction task oriented to pose estimation,a new feature detection network with HRNet as the main network is proposed,and additional semantic branching is introduced to improve the detection accuracy of the system for single-instance or multi-instance pose estimation,and the focal loss is improved to make it more suitable for the pose estimation task.(4)Aiming at the problems that traditional feature suppression methods are difficult to participate in the backpropagation of neural networks and the accuracy loss caused by the integer index of images,a weighted average coordinate method is proposed to suppress features,so that the coordinates obtained by suppression are closer to the real value and the accuracy of the pose estimation results is further improved.In this thesis,the experimental design and verification are based on the Linemod and Occasion-Linemod datasets,and the system performance is evaluated from the three perspectives of ADD(-S),2D Projection and5cm5°,and the ablation experiment is designed to better express the performance improvement of each part of the system.In order to further prove the positive impact of the proposed system on other downstream technologies,this paper designs verification experiments for three aspects:augmented reality,robotic arm grasping and autonomous driving.The final experimental results show that the six-degree-of-freedom pose estimation system proposed in this paper has higher accuracy than traditional methods,and shows better robustness under highly occlusion conditions,which provides strong support and guarantee for various downstream technologies.