Research And Application Of Virtual Reality Fusion Algorithm Based On Pose Estimation

With the application fields of artificial intelligence becoming more widespread,augmented reality technology has drawn more and more attention as an important means of human-computer interaction.However,the pose estimation of camera based on template marker cannot meet the need of complex scenes in the future.In this paper,as shortcomings of the marker-based registration method in augmented reality,with the consideration of the scalability of the application scene and the diversity of interactive means,we use the video sequence information to rebuild the 3D structure information and the timing information of the target scene to estimate the 3D pose of the continuous motion camera.The main contributions include the following aspects:(1)Implement the 3D registration method of visual SLAM augmented reality based on image feature matching.Aiming at static scenes with rich texture scenes,the image features extracted from continuous video sequences are used to match the feature points in different images corresponding to the same scene point,rebuild the three-dimensional structure between adjacent video sequences and estimate camera parameters,and then solve the camera motion trajectory.(2)Propose an augmented reality 3D pose estimation method based on deep neural network in the dynamic scenes.Aiming at complex dynamic scenes with changing motion,we build an end-to-end learning model for input image sequence based on deep neural network,and convolutional neural network is used as a high-level feature extractor.Meanwhile,a recurrent neural network is used to establish the temporal relationship between consecutive video frames,and complete the 3D pose estimation of the continuous motion of the camera.It avoids the situation that the feature of image is not extracted well due to the rapid movement of the camera and the continuous motion change of the scene.