Research And Implementation Of Visual SLAM System For Outdoor AR Applications

There is an urgent need for the augmented reality system in the military,so that soldiers can maintain the state of perception and strengthen the ability to respond to reduce the risk of military action.Visual SLAM is the core technology of augmented reality system,especially for the outdoor complex environment.This paper is focused on more robust visual SLAM.Based on the ORB-SLAM2,this paper improve the Visual Odometry and optimization module,and design a visual SLAM system suitable for outdoor complex environments.It was proved that the improved algorithm and designed system is robust that stable and continuous under harsh outdoor conditions.The main work and contributions are as follows.The basic algorithm and theoretical knowledge of visual SLAM are studied.Compared with the mainstream visual SLAM system,the visual SLAM system framework conforming to the actual outdoor augmented reality applications is selected as the basis for subsequent research.Two excellent visual SLAM systems based on feature point method and direct method are studied that include ORB-SLAM2 and LSD-SLAM.It is found that the visual SLAM based on feature point method is much more suitable for outdoor augmented reality applications.The ORB-SLAM2 system architecture is used as the basis for system design and improvement.The various modules of visual SLAM based on feature point method are studied.Firstly,an improved algorithm of Visual Odometry in the outdoor environment is proposed for the problem that the Visual Odometry has feature extraction speed,weak texture,repeated texture and extremely long polar line.So the feature extraction speed of the Fast algorithm in the known environment is improved,and the repeatability of the extracted features when relocalization is improved.It is also improved the speed of feature matching between consecutive frames that a high-speed mismatch culling scheme based on the combination of GMS and RANSAC.These improvements maintain the real time and significantly increase accuracy of the Visual Odometry.Secondly,for the problem of Optimization module precision,the method of increasing the confidence of feature points in Bundle Adjustment is adopted,which improves the overall accuracy.It is verified system stability and continuity through design experiment.The experimental results show that the Visual Odometry has a little real-time improvement and the robustness is obviously enhanced.The overall system has a long running time,continuous and stable posture,and improved positioning accuracy.