We implemented a use WiFi to facilitate image-based positioning system system to avoid confusion inside the building caused by similar decoration. When the position information is obtained substantially WiFi-based positioning thread, targeting a thread retrieves the image that best matches the image based on, and the camera pose clustering the images associated with a different candidate positions. Select the image cluster closest to the WiFi positioning results in precise camera pose estimation. WiFi use greatly reduce the search area to avoid the situation in the 3D model an extensive search millions of descriptors. In the image-based positioning stage, we propose a novel 2D to 3D to 2D positioning frame that follows the policy from coarse to fine, can quickly locate a plurality of positions in a query image candidates, and performing local feature matching after selecting the correct positioning and image position by the WiFi, the position of the camera can be estimated. The entire system show a significant advantage in combining images and WiFi signals to the localization task, and has great potential for deployment in practical applications.
positioning allowed wifi (a few meters, using methods that do not know), wifi positioning affected by environmental changes, we are also positioned wifi only three degrees of freedom (in position).
Based on the positioning of the image, more robust to environmental change? Six degrees of freedom, grade cm; are not similar to the scene;
Wifi positioning of this paper is not the fingerprint is based on triangulation method
image positioning section, which is to look at the database
AS All at The Training ImagesRF Royalty Free are reconstructed INTO 3D Model through Structure-fromMotion (SfM) Reconstruction Method,, the each Image in at The Database IS offline part associated with the 3D position in the building . The ah, I do now, and very close to the SFM up
The authors do feel very close to SLAM, but is written in image query, may mainly query, to find the last position on 3d points.
Method SfM [5,10,39] 3D point cloud is applied to allow the task to improve the accuracy of positioning. Irschara and so on. [11] retrieved 3D point having the best matching image descriptor. Li et al. [20] through the mutual visibility information matches the 2D image and the 3D model to find a correspondence between 2D, 3D. Simultaneous Localization and Mapping (SLAM) algorithm [4,6,18,38] Online construct 3D models and simultaneously positioned objects. However, SLAM usually accumulate errors in the process of building the model, it is often suitable for small scenes. Moreover, SLAM need computing power and additional equipment, such as laser radar, radar and a stereo camera, for ordinary users, the indoor localization is not ideal. Sattler and so on. [35] proposed the use of visual word matching a 2D image to the 3D model directly, to accelerate the process. Similarly, Liu et. [22] indexing 3D models in the visual vocabulary tree, and retrieve 3D scene from the vocabulary tree for the query image. [8] Embedded Random Ferns 3D point to train a classifier to establish a corresponding relationship query image to the 3D model by the 3D point descriptor is classified as class. [25,26] to delete the features extracted from the foreground by adding the correct correspondence and video applications, thereby enhancing the 3D pointing. Support queries using 3D models of images accurately estimate posture. However, 3D models consists of millions of descriptors. Search the entire model usually spend a lot of time. Using various techniques (e.g., visual word and key frame) to accelerate the matching process. These methods of selecting samples matched to narrow the search to quickly locate. On the other hand, a reduced possibility of finding the best match so that the range of the descriptor is also reduced, by excluding the search area also includes maintaining a high possibility that a correct match the query image.