3D Tracking Technology For Augmented Reality

Augmented reality technology requires accurate understanding of the spatial semantic information of the objective three-dimensional world in order to achieve accurate registration of virtual and real objects.3D tracking aims to continuously estimate the Euclidean transformation relationship between the object and the camera in a time series,that is,to continuously estimate the pose parameters of six degrees of freedom.This technique plays an important role in augmented reality.With the popularity of smart mobile devices and other wearable products,vision-based three-dimensional tracking technology has received much attention in AR research due to its simplicity and precision,and has a wide range of practical application.This paper focuses on three-dimensional object tracking problems based on monocular RGB images.Three-dimensional tracking has been studied for decades,but there are still many challenging factors that limit its application in practice.On the one hand,the physical properties of the object itself vary greatly.On the other hand,the external environment has various tricky factors,such as complex background,motion blur,and rapid displacement.In particular,in the case of monocular vision without spatial geometric prior knowledge,the inference of two-dimensional to three-dimensional information is highly susceptible to interference.In recent years,deep learning has shown strong feature abstraction and model fitting ability,and made great breakthroughs in many a cademic fields.The application of deep learning to 3D tracking problems has gradually developed into a new research trend.The method can learn the feature description of the object pose from the training data,and has the advantage that the traditional method does not have.Therefore,this article combines 3D tracking issues with deep learning and does the following:Firstly,a six-degree-of-freedom pose estimation method based on reconstructed autoencoder is proposed.Combined with the characteristics of pose estimation,the task is reasonably divided into two parts:position prediction and orientation prediction.According to the principle of the denoising autoencoder,the factors such as complex background,motion blur and occlusion between objects are suppressed during the network model learning process,and the main features associated with the target appearance are extracted as an important basis for pose classification.Since the network has a segmentation detection function,the complete position component is restored by parsing the 2D bounding box according to the proportional relationship in the camera imaging model.Secondly,an edge-based pose optimization and verification method is proposed.The pose optimization process uses contour edge features without texture,globally matches 2D observation contours and 3D projection contours,and establishes a least squares problem based on edge distance for nonlinear iterative solution.In order to further verify the accuracy of the pose,the edge gradient isotropic index is proposed as the criterion for pose verification.Finally,implement a video-based automated 3D tracking technology.This technology enables automatic initialization of the tracking process and subsequent tracking of 3D objects in real time.In the initialization state,the system starts the neural network to estimate the rough pose,and uses the reconstructed graph to filter the trusted edges for further pose refinement.In the tracking state,the system is initialized in the previous frame pose,focusing only on the interframe motion,and using the edge distance energy function to perform the pose refinement.In order to ensure the stability of the tracking,the accuracy of the pose estimation is determined according to the edge contour index,and the current tracking state is determined.In summary,this paper implements 3D tracking technology based on deep neural network,and carries out experimental tests on multiple sets of data.The experimental results show that the algorithm has good effects on various performance indicators,and can achieve geometric consistency of virtual and real fusion effect in real time and accurate way which has practical value.