Research On Object Pose Estimation Technology For Augmented Reality

Augmented reality technology is an emerging technology developed in the field of computer vision,graphics,pattern recognition and image processing based on virtual reality.Its purpose is to accurately superimpose computer-generated virtual objects into real scenes and achieve seamless integration of the real and virtual scenes.The realism of augmented reality depends to a large extent on the accuracy of the seamless integration of virtual and real scenes,while the model parameters of the seamless integration are mainly derived from the pose estimation technique.The method of pose estimation based on monocular vision has the advantages of low cost,easy operation and more application scenarios and became the hot point of academic research.Based on the application background of augmented reality,this thesis studies the object pose estimation technology based on monocular vision.The research content is divided into two directions according to the attribute of the marker:object pose estimation technology based on artificial markers and object pose estimation technology based on natural markers.The details are as follows:Aiming at the problem that the markers in ARToolKit are sensitive to illumination and occlusion,this thesis proposes a marker detection algorithm based on LSD algorithm.The LSD algorithm detects the line segments in the image based on the gradient information of the pixels.Based on this,a set of algorithm flow is proposed to detect the region of the marker.Thereby positioning the sub-pixel level key points of the marker.On the basis of this,the original marker recognition algorithm in ARToolKit is improved.It is proved by experiments that the detection and recognition algorithm proposed in this thesis can solve the problems of occlusion and illumination sensitivity of markers in ARToolKit and improve the robustness of the system.The pose data obtained by the ARToolKit algorithm is highly uncertain and the detection algorithm of the marker proposed in this thesis can't work when the occlusion area is too large.In this thesis,a multi-marker pose estimation algorithm based on confidence is proposed.The data obtained from multiple markers are weighted and fused according to the confidence level of each marker.The proposed algorithm solves the problems existing in the original algorithm.After experimental comparison and analysis,the feasibility and effectiveness of the algorithm are proved.For the pose estimation based on natural markers,this thesis mainly research the head pose estimation based on the face of people.The traditional method needs to extract the features of the face and the selection of the feature type and the quality of the feature extraction have a greater impact on the data of the pose.To deal with these problems,this thesis proposes to use CNN network to perform pose estimation of the head and use the method of transfer learning to train our network based on the pre-trained model of VGG16.The trained network can estimate the head pose through the face image,and the accuracy is higher than the traditional method,and the expected effect is achieved.In this thesis,the proposed algorithms are compared with the original algorithms based on artificial markers and natural markers,the proposed algorithms have higher precision and better data stability.They provide high-quality data for the seamless integration of the real and virtual scenes to make the augmented reality effect more realistic.