Citation Context

...onceptually different approach, since exhaustive image comparison or trying all “mountain peaks” simply does not scale to a large-scale geo-localization problem. Similarly, for urban localization, in =-=[13]-=- an upward looking 180◦ field of view fisheye is used for navigation in urban canyons. The authors render untextured city models near the predicted pose and extract contours for comparison with a quer...

Citation Context

...nd the one obtained using Equation 6 (below), accounting for geometric inaccuracies. False changes due to missing details on the building facade disappear in the latter. An alternative is provided by =-=[8, 13, 16]-=-. For each panoramic image, an object class segmentation is performed in order to estimate the building outlines in these images [7]. Let St denote the building outlines estimated on the image It. Eac...

Citation Context

...ner of the intersection), it is limited in urban settings, near buildings, trees and bus stops, where errors due to reflections and signal dropouts prevent accuracy being better than within a few street addresses [6]; thus GPS cannot offer direct assistance in knowing where it is safe to cross. Computer vision is a natural choice of technology to infer the presence and location of crosswalks in the traveler’s environment, since this function nicely complements the information provided by GPS. There is a variety of work (e.g., [16]) on computer vision-based selflocalization, including a system [11] that infers location based on the detailed appearance of the skyline and a realtime system [3] that infers location from panoramic images on mobile devices. However, such systems require the use of detailed 3D models of urban environments, which are complex, memory-intensive and may be difficult to acquire. Instead of relying on a complex 3D model we chose to harness a simple and readily available 2D model of the urban environment based on satellite imagery such as can be obtained freely from Google Maps. In addition to its simplicity, the 2D image model of the intersection has the added adva...

Citation Context

...ed to uncertainties in the order of kilometers. Urban approaches can typically localize much more precisely but they assume presence of certain structures like nearby buildings with known 3D geometry =-=[9]-=- or facades that have a sufficient number of stable local image features on them [5], [10]. Maybe the closest work to ours in the literature is the estimation of viewing direction based on a digital e...

Citation Context

...ifferent approach, since exhaustive image comparison or trying all possible “mountain peaks” simply does not scale up to a large-scale geo-localization problems. Similarly, for urban localization, in =-=[27]-=- an upward looking 180◦ field-of-view fisheye is used for navigation in urban canyons. They render untextured city models near the predicted pose and extract contours for comparison with the query ima...

Citation Context

...ose generated from candidate camera locations. This approach has been shown to be very robust and reasonably accurate, over image sequences corresponding to camera displacements of hundreds of meters =-=[40]-=-. 6.5.2. Generic self-calibration approach In [19], we have finalized and improved upon previous works on self-calibration for generic camera models. From 2004 on, we performed research on generic cam...

Citation Context

...ilar approach for use in urban canyons where skyline imagery of buildings are processed to extract features such as edges of buildings, which are compared to aerial or satellite maps of the same area =-=[38]-=-. Accuracies of 2 m were achieved but the algorithm has several limitations mainly that skylines of buildings are expected to be close to the camera. Another limitation is the fact that nighttime oper...