Research On Object Detection And Tracking Technology Oriented To Augmented Reality

Augmented reality is a technology based on 3-Dimension(3D)Registration,virtual-real fusion,and human-computer interaction technology,whose goal is to implement the fusion of virtual objects and real scenes.The core problem of the augmented reality system is to recognize and track objects in the real environment accurately and quickly,and then determine the current position and posture of the object.Therefore,the virtual object can be accurately placed in the corresponding position.Therefore,this paper focuses on the research of 3D registration technology in AR.It mainly uses deep learning-based methods to research and improve the object detection and pose estimation method,and proposes an object recognition system that is robust and can meet real-time needs.The main contributions of this paper are as follows:Aiming at the problem of low detection accuracy of traditional object detection models under object occlusion and complex background conditions,an object detection algorithm based on a self-attention mechanism is proposed.This paper improves the backbone feature extraction network structure of the Retina Net object detection model.The self-attention mechanism is introduced to improve the ability of the network model to extract and learn image features,which effectively improves the detection accuracy and robustness of the model.In order to ensure that the model improves the accuracy and high detection efficiency,the differential evolution algorithm is applied to optimize the Anchor size in the dataset.We conducted a comparative experiment on the PASCAL VOC dataset.The experimental results show that our proposed model has better robustness in the real-world image that may have the situation of object occlusion and complex background.In order to realize the pose estimation of 3D objects,an end-to-end 6D pose estimation algorithm is implemented.6-Dimension(6D)poses are used to represent the poses of 3D objects in world coordinate.In order to implement the task of detecting 3D objects,we have improved the structure of the original object detection model.Firstly,the 2D mapping point coordinates of the eight corner points and one center point of the 3D bounding box describing the object are obtained through the convolutional neural network in the two-dimensional image.First,the convolutional neural network is used to extract the 2D coordinates of the object's 3D edge detection box,which includes eight corner points and a center point.After obtaining the mapping relationship between 3D and 2D points,the EPnp algorithm is applied to obtain the translation component and rotation component of the object in the 3D coordinates to implement the pose estimation of the 3D object.This paper conducts a comparative experiment with the existing pose estimation algorithm on the Line Mod dataset.Experimental results show that the improved pose estimation model can reach the state-of-the-art algorithms in both 2D reprojection error and ADD.In summary,this paper has not only realized the detection of 2D and 3D objects formed by RGB image format,but also successfully applied to the augmented reality system.The experimental results show that the augmented reality-oriented object detection system proposed in this paper,which has higher detection accuracy and practical application value.