Visual Positioning with Image Database and Range Camera

W-W. Kao, I-J. Chu

Abstract:	There has been some works on smartphone or vehicle navigation using geocoded image database. In these works, images taken at various locations in a particular environment are recorded with their photo-taken positions and compiled to form a database. Place-recognition is then achieved by associating the real-time image taken from unknown location in the environment with a geocoded image from the database that has the most similar appearance. Various algorithms are developed to associate images with different perspectives.by their content appearance. Position of the real-time image can then be deduced from the geocode information of the associated database image, often with very rough accuracy. In this research, prior geocode information is not required for images in the database. However, stereo camera is used to acquire real-time image for positioning purpose. The additional depth information from stereo vision allows proper feature measurements for images with different perspective and enables more accurate position calculation subsequently. In the place-recognition phase, image with similar scene objects as real-time image can be retrieved from the database through appearance-based algorithms. With common feature image coordinate and depth information provided by stereo vision, camera view angle and position differences between the real-time image and the reference database image can be derived. By taking the current image and the reference image photo-taken position as unknown states and formulate the problem in typical SLAM fashion, both the camera position and reference image position can be estimated using the image measurements. Consequently, non-geocoded images in the database can be updated with their corresponding geocode information for future usage. Additionally, positioning accuracy of the camera can be greatly improved. Indoor positioning of an unmanned robot vehicle equipped with a smartphone with 3-D stereo camera is performed. In the experiment the robot first explores an unknown environment by taking photos at random locations to establish image database. The robot then re-visits the environment and place-recognition is performed to extract image from the database that has the most similar appearance compared with the real-time camera image. Once this is achieved, SLAM process starts to estimate robot own position and the database image position. Experimental results will be presented to demonstrate the performance of the proposed algorithm.
Published in:	Proceedings of the 26th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2013) September 16 - 20, 2013 Nashville Convention Center, Nashville, Tennessee Nashville, TN
Pages:	461 - 466
Cite this article:	Kao, W-W., Chu, I-J., "Visual Positioning with Image Database and Range Camera," Proceedings of the 26th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2013), Nashville, TN, September 2013, pp. 461-466.
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In