Contemporary smart phones use various localization methods based on GPS, cellular networks and Wifi networks. However, none of the methods available today is able to reliably and accurately determine a user's location inside buildings.
Normally, no infrastructure supporting localization is available inside buildings. Similarly, smartphones may not be equipped with specialized localization hardware.
With recent advances in content based image retrieval (CBIR), fast visual localization of mobile devices becomes feasible. Accordingly, the visual information that may be made available through a phone's camera is used for location estimation. By comparing the features visible in the image taken by the camera to geo-tagged reference images recorded previously during a mapping run, the location of the camera can be determined.
Utilizing video recordings of a mobile device as a visual fingerprint of the environment and matching them to a geo-referenced database provides pose information in a very natural way. Hence, location based services (LBS) can be provided without complex infrastructure in areas where the accuracy and availability of GPS is limited. This is particularly interesting for indoor environments, where traditional localization methods like GPS are unavailable.
However, the application of CBIR to mobile location recognition implies several challenges. The complex 3D shape of the environment results in occlusions, overlaps, shadows, reflections, etc., which require a robust description of the scene. Bag-of-Features based image representations are able to fulfill these requirements, however they require a huge amount of reference images in order to be useful for localization.
Vision-based localization systems make use of local image features, organized in a searchable index using content-based image retrieval (CBIR) methods. Once trained on a set of reference images, CBIR systems are able to rapidly identify images similar in appearance to a query image. However, when applied to the problem of visual localization, two major problems surface:
Limited accuracy: In order to provide reference images for the image retrieval system, the environment needs to be mapped, i.e. images have to be captured at various locations and orientations, and corresponding map coordinates have to be stored. This is commonly achieved by mapping trolleys which automatically capture images and acquire a 3D point cloud model as it is moved through the environment. Although automated to a large degree, mapping buildings on a large scale is a time-consuming and tedious endeavour, and it is impossible to capture images at every combination of location and orientation that might occur during localization. In practice, images are captured along a single trajectory only, drastically limiting the resolution of position and orientation estimates as returned by the image retrieval process.
Perspective distortion: The limited affine and perspective invariance of feature descriptors is a severe problem, as a location can be recognized only if a reference image with a pose similar enough to the query image exists. There has been extensive work on improving the robustness of feature descriptors under perspective distortion. However, robustness is gained at the expense of distinctiveness, hence such approaches tend to increase recall only, but not precision.
It is known to apply content based image retrieval approaches for location recognition in textured outdoor environments [1, 2, 10, 11]. Indoor environments, however, are more challenging, as only few distinctive features are available and perspective distortion is more pronounced, especially in narrow corridors.
Attempts to address perspective distortions are described in  and . However, these methods are computationally expensive or do not have to deal with complex geometric variations.
Further, it is known to determine information on the 3D structure of an environment, e.g. via laser scans, and to use such information to generate locally orthogonal projections. In  there is described a combination of conventional, perspective images with orthogonal projections of building facades to increase invariance with respect to the viewpoint. Increasing feature invariance however, generally deteriorates distinctiveness, which is particularly unfavourable in texture-poor indoor environments.
From  it is known to generate viewpoint invariant patches (VIP) to improve robustness in respect of 3D camera motion.
The generation of synthetic views is described in . However, the approach described in this document may be insufficient in case of sparse reference imagery. Further, occlusions are not handled by this approach, which is of particular importance in indoor environments where obstacles and walls restrict visibility.
From  it is known to generate orthogonal projections of buildings facades. Query images are normalized to surface-parallel views after analyzing them for vanishing points. However, this approach too is expensive in terms of processing.
List of References
-  G. Baatz, K. Koser, D. Chen, R. Grzeszczuk, and M. Pollefeys. Leveraging 3D city models for rotation invariant place-of-interest recognition. International Journal of Computer Vision, 96(3):315--334, Feb. 2012.
-  D. Chen, G. Baatz, K. Koser, S. Tsai, R. Vedantham, T. Pylvanainen, K. Roimela, X. Chen, J. Bach, M. Pollefeys, B. Girod, and R. Grzeszczuk. City-scale landmark identification on mobile devices. In CVPR, pages 737--744, Colorado Springs, USA, June 2011.
-  D. Chen, S. S. Tsai, V. Chandrasekhar, G. Takacs, J. Singh, and B. Girod. Robust image retrieval using multiview scalable vocabulary trees. In Proc. of SPIE, number 1, pages 72570V--9, San Jose, USA, Jan. 2009.
-  R. Huitl, G. Schroth, S. Hilsenbeck, F. Schweiger, and E. Steinbach. TUMindoor: An extensive image and point cloud dataset for visual indoor localization and mapping. In IEEE International Conference on Image Processing, Miami, USA, Sept. 2012.
-  A. Irschara, C. Zach, J.-M. Frahm, and H. Bischof. From structure-from-motion point clouds to fast location recognition. In CVPR, pages 2599--2606, Miami, USA, June 2009.
-  T. Liu, M. Carlberg, G. Chen, J. Chen, J. Kua, and A. Zakhor. Indoor localization and visualization using a human-operated backpack system. In Indoor Positioning and Indoor Navigation (IPIN), pages 1-10, Sept. 2010.
-  J.-M. Morel and G. Yu. Asift: A new framework for fully affine invariant image comparison. SIAM Journal on Imaging Sciences, 2(2):438--469, Apr. 2009.
-  D. Nistér and H. Stewénius. Scalable recognition with a vocabulary tree. In CVPR, pages 2161--2168, New York, USA, June 2006.
-  R.B. Rusu and S. Cousins. 3D is here: Point Cloud Library (PCL). In IEEE International Conference on Robotics and Automation (ICRA), Shanghai, China, May 9-13 2011.
-  G. Schroth, R. Huitl, M. Abu-Alqumsan, F. Schweiger, and E. Steinbach. Exploiting prior knowledge in mobile visual location recognition. In IEEE ICASSP, Kyoto, Japan, Mar. 2012.
-  G. Schroth, R. Huitl, D. Chen, M. Abu-Alqumsan, A. Al-Nuaimi, and E. Steinbach. Mobile visual location recognition. IEEE Signal Processing Magazine, 28(4):77--89, July 2011.
-  J. Sivic and A. Zisserman. Video Google: A text retrieval approach to object matching in videos. In International Conference on Computer Vision, volume 2, pages 1470--1477, Beijing, Oct. 2003.
-  C. Wu, B. Clipp, X. Li, J.-M. Frahm, and M. Pollefeys. 3D model matching with viewpoint-invariant patches (VIP). In CVPR, Anchorage, USA, June 2008.
The present invention aims to address the above problems. In particular, but not exclusively, the present invention aims to provide a robust method of visual localisation that is suitable for indoor environments. Moreover, the present invention aims to provide a method of visual localisation that can provide satisfactory results on the basis of a relatively low number of reference images.
Summary of the invention
According to an aspect of the present invention, there is provided a method of visual localisation, comprising: generating a plurality of virtual views, wherein each of the virtual views is associated with a location; obtaining a query image; determining the location where the query image was obtained on the basis of a comparison of the query image with said virtual views.
In particular, in an embodiment of the invention there is provided a method for visual localization based on pre-computed virtual images or image patches at arbitrary, user-defined locations from sparse reference imagery. The method relies on knowledge about planar regions present in the scene, and projectively transforms reference images to the virtual view's location. Instead of using planar models, more sophisticated three-dimensional models of the environment can be constructed in order to generate virtual views for localization.
Query images are matched to virtual views using suitable image retrieval techniques. In exemplary implementations, the invention provides a localization system that is able to robustly determine the orientation and position on a meter-level within fractions of a second.
Further, in an embodiment there is provided a method for locating images depicting the parts of a plane that are visible in a virtual view. There is also provided a localization system based on this approach that is able to determine the position and orientation of images at locations where no reference images have been captured.
The present invention is able to use sparsely distributed reference images to perform a reliable visual localisation. In an embodiment, local image features are extracted from virtual viewpoints by identifying planar regions in the virtual images and applying corresponding homography transformations to reference images. By extracting local features from the generated image patches, and combining features from all patches for a viewpoint into a document that is indexed by the CBIR system, the present invention enables determining camera poses far off the mapper trajectory solely based on image retrieval results.
The present invention which is based on visual information is particularly suited for indoor localization.
In particular, the present invention provides a method of visually localising a mobile device (or in more generic terms, any image) by determining the visual similarity between the image recorded at the position to be determined and localized (virtual) reference images stored in a database.
In a preferred embodiment of the present invention, the localised (real) reference images, which have been recorded along a mapping trajectory, are transformed into virtual viewpoints of arbitrary camera positions and orientations in the environment. The viewpoints are represented by their respective bag-of-features vectors and image retrieval techniques are applied to determine the most likely pose of query images. This can be performed at relatively low computational complexity. As virtual image locations and orientations are decoupled from actual image locations, the system is able to wok with sparse reference imagery and copes well with perspective distortion. Experiments have confirmed that pose retrieval performance is significantly increased compared to conventional systems relying on (real) reference images only.
The method according the present invention may be used in connection with a reference database of virtual views that stores the appearance at distinct locations and orientations in an environment, and an image retrieval engine that allows lookups in this database by using images as a query.
The generation of the reference database is an offline process performed after an environment has been mapped. During the mapping phase, may be captured and tagged with their 6-DOF pose (location and orientation) and, in addition, a three-dimensional point cloud model may be acquired. A two-dimensional occupancy grid map may be used as a reference coordinate system and to determine valid locations for virtual views.
The present invention resides in particular in using pre-computed virtual views of the environment, generated from relatively few reference images, for visual localization in a content-based image retrieval (CBIR) scheme. In the following, exemplary embodiments of the invention are described. These embodiments rely on a detection of planes in the scene/environment and are applicable even in situations where no accurate 3D model is available.
Brief description of the drawings
Figure 1 illustrates images captured during mapping and the virtual viewpoints created. At each location, 16 views are computed.
Figure 2 illustrates how the point cloud acquired during mapping is segmented into planes that provide the models for projective transformations.
Figure 3 (a) illustrates how the point cloud from a virtual viewpoint is used to lookup visible planes in a pre-computed point-to-plane map.
Figure 3 (b) illustrates how images are assigned an warped to each plane.
Figure 3 (c) illustrates how the mask keeps track of unassigned plane parts.
Figure 4 illustrates warped image patches for the plane in Figure 3 (c).
Figure 5 illustrates top-ranked query results for four query images, wherein the black square is the ground truth pose. The location and orientation are drawn as circles. No post processing of image retrieval results has been applied.
Description of exemplary embodiments of the invention
As described above, the present invention enables the rendering of (partial) images from arbitrary viewpoints in a 3D scene or environment. In an embodiment of the present invention, in order to simplify the mapping phase and the rendering of novel views, triangulation of points to meshes is avoided, and instead predetermined geometric models, such as planes, are used to represent portions/regions of the environment, e.g. building interiors. As projections of a plane into the image space of two cameras are related by a homography (projective transform), viewpoint changes from one camera towards the other can be simulated by applying the projective transform to the former camera's image. In this embodiment, a reference view is chosen as the former camera and its image is transformed to the virtual camera's view by applying the projective transform, which is a function of pose and calibration of the two cameras (reference view and virtual view) and the plane's position in space. This simplifies the computation of new views from existing images.
In an initial step, a point cloud is acquired, for example by laser-scanning of the environment. The point cloud is segmented into planes. These planes provide the model for projective transformations. Fig. 2
In particular, planes in the point cloud model are identified by fitting horizontal planes (floors and ceilings) and vertical planes (walls) using a sample consensus method. Thereafter, a mapping M of 3D points to plane identifiers is performed. Subsequently, for each point P in the segmented cloud, the set of reference images IP
that depict the given point are determined by, for each reference view, checking whether the point P lies inside the viewing frustum of the reference view, i.e., whether it is contained in the volume that's depicted by the camera. Casting rays from the point towards the respective reference view's camera centre is used to detect occlusions.
Identification of visible planes
In an embodiment of the invention, first, the major planes visible in the virtual view (see Fig. 3a) are determined by casting rays from the centre of the virtual camera through pixels in its image plane into the scene. When a ray hits a scene point, the map M is used to lookup the plane's position in space (the plane parameters). This step is performed for all pixels of the virtual view (although spatial sub-sampling in the pixel domain can be used for efficiency), and the resulting list of planes is sorted by the number of pixels that belong to each plane. For each plane, the algorithm keeps track of the pixels that are part of the plane (see Fig. 3b).
At this point, each plane is processed separately in order to find the reference images with a good view on the 3D points associated with that plane. In one embodiment, the algorithm combines the image lists IP
for all plane points into a single list and applies histogram binning to determine the reference image which covers the plane best. In the following step, this image is warped to the virtual viewpoint and its pixels are removed from the current plane's pixel mask (see Fig. 3c). The image assignment process is repeated until the number of pixels remaining falls below a threshold or no more reference images are available for the plane.
The correct selection of reference images enhances the results. In an embodiment, two constraints are added to the image selection algorithm. First, an upper limit on the angle between the reference image's normal and the plane normal avoids using low-resolution views of a plane. Second, when multiple reference images cover approximately the same number of plane pixels, the one closest to the virtual view's location. This avoids low resolution warping results and prefers reference images with similar perspective.
Image warping and feature extraction
The camera pose of the reference image is denoted by a homogenous 4x4 matrix Tref
, the pose of the virtual image is denoted by Tvirt
. The relative transformation between both views follows as
With a plane defined in Hessian normal form xT
the distance between the plane and the reference image is
The homography H
relating coordinates in the reference image to coordinates in the virtual image is then given by
are the camera calibration matrices for the reference image and the virtual image, respectively.
Using equation 3, the reference image is warped to the virtual viewpoint and local image features are extracted from the resulting image patch (see Fig. 4). For any non-trivial scene, the generated patch contains areas where the plane-to-plane homography is inadequate to express viewpoint change. For this reason, all features outside the pixel mask (see above) are discarded.
Finally, the features extracted from all the planes in a virtual view are combined into a single bag-of-features vector that is indexed by a CBIR system for retrieval during localization.
With the reference database prepared as described above, finding the position as well as the orientation of a camera is achieved by extracting features from the query image and retrieving the most similar virtual views from the CBIR database. This step can be performed very quickly using an inverted index and has been shown to scale well up to millions of documents.
Illustrative implementation of an embodiment
In an illustrative, non-limiting implementation and evaluation of the invention, a dataset containing more than 40,000 images of the corridors and halls of a public building is used. For the evaluation a subset of 3,146 high-resolution close-ups is used, captured along a trajectory of more than one kilometer. The area shown in Fig. 1 is a small portion of this subset.
A simple scheme is used to determine locations where virtual views are created. The floorplan is sub-sampled to a resolution of one meter per pixel, and a virtual location is created for each "free" pixel. The height of the virtual camera is fixed a 1.50 m above ground. To simulate different orientations, virtual views are generated for yaw angles advancing in steps of π/8, creating 16 views per location. In total 6,352 locations and 101,632 views are obtained.
The system is scalable to a considerable higher number of views (e.g. up to 10 million views or more), as is common for conventional visual localisation systems. However, the present invention can provide the same accuracy as conventional systems on the basis of a relatively lower number of views.
The image retrieval system is trained on 24.8 million SIFT features extracted from the image patches for the virtual views (see Fig. 4). An approximate k-means (AKM) quantizer with a vocabulary size of 200,000 visual words and TF-IDF weighting is used. The query time per image on a single thread is around 200 ms, however AKM can easily be configured to perform considerably faster.
The system is queried using images captured at various locations in the mapped environment. To demonstrate that the system is capable of inferring the appearance at arbitrary locations, attention is paid to keeping a distance to the mapper trajectory. Four query images and the corresponding results are shown in Fig. 5. The proposed method is able to robustly determine the correct orientation (quantized to intervals of π/8). The top-ranked image retrieval results concentrate in the immediate vicinity of the query location in almost all cases, only the rank-2 location result for the lower-right query image is a bit off. Nevertheless, the orientation is determined correctly in all cases.
The virtual camera uses the same calibration matrix as the query camera to ensure that the virtual views match what the query camera would see. If the field of view (FOV) between the cameras differs too much, a localization error along the camera's z-axis can occur. For the application of smartphone localization, it can be assumed that the FOVs do not vary considerably between different phone models. Further, the focal length of query cameras may be artificially lengthened simply by cropping the region of interest for feature extraction.
Table 1 shows the mean precision over 252 queries (six frames at 42 locations) achieved by the first result, by the top-3 results, and by the top-5 results, respectively. A precision of 1.0 is achieved if all top-ranked results are relevant. Clearly, the virtual view approach outperforms the conventional approach of using unprocessed reference images. In 56% of all cases, the top-ranked result is a correct location with our virtual view approach, compared to 33% when only reference images are used.
Table 1: Mean precision at cutoff ranks 1, 3 and 5. Relevant views are within radius r around the query location
| ||P @ 1||P @ 3||P @ 5|
|Reference Views (r = 5m)
|Virtual Views (r = 3m)
|Virtual Views (r = 3m)
It will also be appreciated that the above described embodiments are described as examples only, and that modifications to these embodiments are included within the scope of the appended claims.
The present invention is further defined by way of the following non-limiting examples:
According to example 1, the invention is defined by a method of visual localisation, comprising: generating a plurality of virtual views, wherein each of the virtual views is associated with a location; obtaining at least one query image; determining the location where the query image was obtained on the basis of a comparison of the query image with said virtual views.
According to example 2, the method of example 1 further comprises: generating said virtual views from a model of an environment, for example the inside of a building, in which said visual localisation is to be performed; or generating said virtual views from one or more reference images of an environment, for example the inside of a building, in which said visual localisation is to be performed.
According to example 3, the method of example 2 further comprises: extracting one or more features from one of said reference images; and transforming said one or more features to obtain a visual representation of said features at one or more locations different from the location at which said one reference image was generated.
According to example 4, the method of examples 2 or 3 comprises: transforming one or more of said reference images to the location of one or more of said virtual views.
According to example 5, said query image is generated by a mobile device, the method further comprises determining the pose of the mobile device on the basis of said comparison of the query image with said virtual views.
According to example 6, the method of any preceding example comprises: determining one or more planes within said virtual views; and allocating points of said query image to one or more of said planes.
According to example 7, the method of any preceding example comprises: determining a plurality of points in an environment; identify one or more planes in said environment from said plurality of points; and mapping some or all of said points to said planes.
According to example 8, the method of example 7 comprises: for each point that has been mapped to a plane, determining one or more reference pictures that include said point.
According to example 9, the method of examples 7 or 8 comprises: determining the trajectory of rays between one or more of said points and a given virtual viewpoint, thereby to detect planes within the virtual view from the given virtual viewpoint, and preferably repeating this step for each pixel associated with the virtual view from the given virtual viewpoint, and preferably sorting the planes by the number of pixels that belong to each plane.
According to example 10, the method of examples 7 to 9 comprises: processing each of the detected planes to determine reference images containing points that have been mapped to the detected planes, and assigning detected planes to reference images, in particular assigning that reference image that best matches a given plane; and optionally repeating said assigning for some or all of the detected planes, in particular until the number of pixels associated with non-assigned planes reaches a predetermined threshold or there are no more unmapped reference images.
According to example 11, the method of example 10 comprises: warping a reference image, in particular the reference image that best matches said given plane, to a given virtual viewpoint, thereby to generate a virtual view from said viewpoint.
According to example 12, the method of example 10 or 11 comprises: wherein said assigning is performed only if the angle between the normal of the reference image and the normal of a given detected plane is below a predetermined angle, and/or wherein, if there is more than one candidate reference image for assigning, the one of the candidate reference images closest to the virtual viewpoint is selected.
According to example 13, the method of any preceding example comprises: extracting features from said virtual views, and preferably combining extracted features into a single bag-of-features vector that is indexed by a CBIR system for retrieval during localization.
According to example 14, the method of example 13 comprises extracting features from said query image and retrieving the most similar virtual views from the CBIR system, thereby to determine the location and the orientation of a camera used to generate the query image, wherein the retrieving step is preferably performed by using inverted files.
As example 15, a possible embodiment of the invention is an article storing instructions that, when executed by a data processing system, cause the method of any preceding examples to be performed.
A method of visual localisation, comprising:
recording localised reference images of an environment along a mapping trajectory;
determining information on the 3D structure of the environment,
detecting planes in the environment;
generating a plurality of new views, wherein each of the new views is associated with a location and orientation, comprising:
using the planes to represent portions or regions of the environment; and
using the planes in computing the new views from the reference images, wherein a viewpoint change from a first camera view corresponding to a reference view to a second camera view corresponding to a new view is simulated by applying a projective transformation to a reference image of the first camera, wherein the projective transformation is a function of pose and calibration of the first and second cameras and plane position in space and the projection of a plane from the reference view to the new view is represented by a homography;
obtaining at least one query image; and
determining the location where the query image was obtained on the basis of a comparison of the query image with said new views.
2. The method of claim 1, wherein said query image is generated by a mobile device, the method further comprising determining the pose of the mobile device on the basis of said comparison of the query image with said new views.
The method of any preceding claim, comprising:
determining a plurality of points in an environment;
identifying one or more planes in said environment from said plurality of points; and
mapping some or all of said points to said planes.
4. The method of claim 3, comprising:
for each point that has been mapped to a plane, determining one or more reference pictures that include said point.
5. The method of claim 3 or 4, comprising:
determining the trajectory of rays between one or more of said points and a given viewpoint, thereby to detect planes within the new view from the given viewpoint, and preferably repeating this step for each pixel associated with the new view from the given viewpoint, and preferably sorting the planes by the number of pixels that belong to each plane.
The method of any of claims 3 to 5, comprising:
processing each of the detected planes to determine reference images containing points that have been mapped to the detected planes, and assigning detected planes to reference images, in particular assigning that reference image that best matches a given plane; and
optionally repeating said assigning for some or all of the detected planes, in particular until the number of pixels associated with non-assigned planes reaches a predetermined threshold or there are no more unmapped reference images.
7. The method of claim 6, further comprising:
warping a reference image, in particular the reference image that best matches said given plane, to a given viewpoint, thereby to generate a new view from said viewpoint.
8. The method of any of claim 6 or 7, wherein said assigning is performed only if the angle between the normal of the reference image and the normal of a given detected plane is below a predetermined angle, and/or wherein, if there is more than one candidate reference image for assigning, the one of the candidate reference images closest to the viewpoint is selected.
9. The method of any preceding claim, comprising extracting features from said new views, and preferably combining extracted features into a single bag-of-features vector that is indexed by a CBIR system for retrieval during localization.
10. The method of claim 9, comprising extracting features from said query image and retrieving the most similar new views from the CBIR system, thereby to determine the location and the orientation of a camera used to generate the query image, wherein the retrieving step is preferably performed by using inverted files.
11. An article storing instructions that, when executed by a data processing system, cause the method of any preceding claim to be performed.