In Section 2, existing works on vision-based localization are reviewed, with focus on appearance-based techniques. Section 3 introduces polar higher-order local auto-correlation (PHLAC) which is then used for probabilistic localization based on Sequential Monte Carlo, described in Section 4. Results from experiments using a real robotic system are then presented in Section 5, along with an analysis of the PHLAC vectors and their robustness against noise and occlusion. Section6 deliberates on how the extracted vectors can be made more distinct for different locations and how invariance against illumination can be introduced directly into the extraction process. Finally, in Section 7, main contributions and results are summarized, and major strengths and weaknesses of our current system are discussed. 2. Vision-based localization A camera-based equivalent of a typical distance-based localization system would be acquiring a detailed threedimensional model of the environment [39]. This 3D model can then be used during localization to generate the expected 2D projections (camera images) at different locations. Instead of being able to internally generate complete camera images, the system could settle for being able to predict what features [49] would be detected at each location. The features in this case make up a sort of sparse 3D map of the environment. However, the 3D model-based approaches regularly depend on auxiliary distance sensors like stereo [8] and trinocular [10] cameras. An alternative to the Cartesian 3D models is to use an appearance-based approach [22]. Appearance-based localization, in its simplest form, involves taking raw snapshots at various locations and storing them along with positioning information. The current camera image can then subsequently be matched against these memorized images to find the most probable current location. お願いします＾＾；
