An Augmented Reality System Based On Monocular Visual Odometry

Augmented reality technology is a new type of human-computer interaction technology that superimposes virtual information “seamlessly” in real scenes,which has a strong sense of immersion and realism.Currently,augmented reality technology can be mainly divided into methods based on two-dimensional image recognition and three-dimensional methods including the ones based on Visual Odometry algorithm.The latter one has become the trend of augmented reality technology as it can achieve better interaction performance and greater practicality.Visual Odometry algorithm is the algorithm that estimates the pose of a moving object and simultaneously predicts the three-dimensional structure of the surrounding environment.Semi-direct method is a kind of Visual Odometry algorithm that estimates the pose by minimizing the photometric error of the sparse features.This method has the advantage that its running speed is very fast,but its accuracy of pose estimation is not very high,compared to other Visual Odometry algorithms.This paper first analyzes the algorithmic process of the front-end model and the back-end model of the semi-direct method's representative work SVO: Fast Semi-Direct Monocular Visual Odometry(referred to as SVO)in detail.In order to improve the robustness and the accuracy of the pose estimation of SVO algorithm,based on SVO algorithm,the following work has been done:1.For the problem that SVO algorithm is unstable or the pose estimation may fail due to the large difference of gray level between temporally adjacent image frames,the improvement of gray level consistency based on histogram equalization is used in this paper,and it effectively reduces the difference of gray level between temporally adjacent image frames and improves the stability and accuracy of the algorithm.2.For the problem that the SVO algorithm estimates poses inaccurately of the moving camera when the camera is moving quickly,this paper introduces an initialization based on motion prior to improve the sparse image alignment step and also improves the accuracy of pose estimation when the camera is moving quickly.3.For the problem that is caused by frequently adding keyframes in SVO,this paper proposes a selective keyframe culling strategy for updating the map,which can effective improve the accuracy of the pose estimation.4.Aiming at the problem that virtual objects in augmented reality systems usually are drawn according to some reference plane,this paper introduces a plane estimation algorithm based on RANSAC in the augmented reality module,to estimate the possible parameters of the three-dimensional plane in the scene.Finally,this paper combines the proposed improved SVO algorithm which is used as a camera pose estimation module and the augmented reality module,to accomplish a complete augmented reality system.