| Summary: | 碩士 === 國立交通大學 === 資訊科學與工程研究所 === 106 === With the rapid growth in mobile devices and wearable cameras, vision-based positioning has become a hot research topic. Many multimedia and robotic applications such as augmented reality, autopilot, and mobile robots, rely on highly-accurate localization results. In this paper, we propose a real-time and drift-free vision-based global positioning system (VB-GPS), which combines two state-of-the-art localization approaches, (i.e., visual simultaneously localization and mapping (SLAM) and model-based localization with a pre-trained 3D point cloud model), together to take advantages of both methods. Visual SLAM can run in real-time and works well even in an unknown environment, but the drift problem is a major issue. Thus visual SLAM cannot be used in multi-user applications. Alternatively, model-based localization provides global positioning without any drift error, but it is time-consuming, and unstable if the scene changes or if any part of scenes is not learned in advance. In addition, because of the estimated global positioning results, the proposed method can be applied to a multi-user scenario and allows multiple users to interact with each other. In the experiments, we compare our method with three state-of-the-art localization methods in six scenes and three scenarios. Results show the proposed method considerably outperforms the others and achieves positioning accuracy with a median error of approximately 0.3 m and 0.95°, even when the camera rotates rapidly, the illumination of the environment is quite different from the pre-trained environment, or when parts of the environment had never learned. These are the main issues for vision-based positioning approaches that normally occur in real world scenarios, but none of the existing methods can be applied to all of them, to the best of our knowledge.
|
|---|
