Research On The Key Problems Of Tracking And Registration Of Moving Target And Scenes For Real-time Augmented Reality

Augmented Reality technology is a hot topic of computer graphics in recent years. Tracking and registration determine the performance of the augmented reality system. Accuracy of tracking and registration directly affect the user experience. Therefore, to improve the accuracy of tracking and registration has important significance. This paper carries out the deep research on tracking and registration of moving target, tracking and registration of blur moving target, tracking and registration of known scene, tracking and registration of unknown scene. The specific contents are as follows:(1) The problems of tracking and registration of moving target are studied. First, this paper presents an improved method of FREAK feature descriptor, and this descriptor is based on the consideration of the spatial structure of the descriptor sampling model on the basis of FREAK descriptors. Each receptive field of IFREAK descriptors stores a plurality of bits, bits depending on the parameter. Experimental results show that compared with FREAK descriptor IFREAK descriptor improves the accuracy of feature points matching. Secondly, based on IFREAK feature descriptor, an improved augmented reality algorithm of tracking and registration of moving targets is proposed. Experiments show that the tracking and registered of moving target is satisfactory.(2) The problems of tracking and registration of blur moving target are studied. As severe motion blur will lead to the failure of target tracking, we propose an ESM-based template matching target tracking algorithm. First, an image structure model of motion-blurred template matching is proposed. Then the ESM algorithm is introduced and motion template is tracked based on ESM-MB. Subsequently, camera shutter estimated time is introduced as a parameter and self-adaptive ESM-MB-ST tracking algorithm of motion blur caused by different shutter speed was proposed. Second, we propose a three-dimensional motion blur rendering virtual objects algorithm. We use the same Image formation model to generate blur. We produce motion blur for 3D motion by mixing 2D distorted and 3D rendering. Finally, an algorithm of tracking and registration of blur moving target based on ESM-MB-ST is proposed. Experimental results show that the tracking and registered of blur moving target is satisfactory and this algorithm is valid.(3) The problems of tracking and registration of known scene are studied. First, because the speed of tracking and registration on the wide area of known scenarios of augmented reality is slow, a classifier based on embedded CCA fern (EC fern classifier) is proposed. The classifier uses a supervised dimensionality reduction method. Canonical correlation analysis (CCA) is introducted. Through training decide the optimum parameters of the binary separation. Second, an algorithm of tracking and registration of known scene based on EC fern classifier is proposed. The algorithm has the offline training section and the online tracking section. Offline training module 3D point cloud is builded on the known scenes. Then feature descriptors of 3D feature point are extracted and EC fern classifier is built. Input video is preprocessed on online tracking module. Feature points of video frames are detected and the detected feature points are described. The extracted descriptors are classified with the established EC fern classifier, namely 2D-3D feature point matching. The epipolar geometry of two images is estimated with PROSAC algorithm. Camera pose is computed and virtual objects are rendered and registered. Experimental show that, the proposed algorithm is a kind of more robust tracking and registration algorithm. The effect of tracking and registered of known scene is satisfactory.(4) The problems of tracking and registration of unknown scene are studied. For tracking and registration algorithm of unknown scene based on the EKF-SLAM, the nonlinear algorithm EKF-SLAM causes error accumulation. And this leads to the estimation results inconsistencies. To solve this problem, an improved extended Kalman filter SLAM algorithm with polynomial is proposed. Use sine and cosine of the angle space as the spatial relationships of the pose of acquisition device to feature. Further nonlinear equations are converted into polynomial equations. Experimental show that, the consistency of the proposed IEKF-SLAM algorithm is better than EKF-SLAM algorithm. On this basis, a tracking and registration algorithm of augmented reality on an unknown scene is proposed. The algorithm has two parallel modules, map building and updating, tracking and registered. Experiments show that the effect of tracking and registered of unknown scene is satisfactory.