A Plane Tracker for AEC-automation Applications

Transcription

1 A Plane Tracker for AEC-automation Applications Chen Feng *, an Vineet R. Kamat Department of Civil an Environmental Engineering, University of Michigan, Ann Arbor, USA * Corresponing author (cforrest@umich.eu) Purpose We propose a new registration algorithm an computing framework, the keg tracker, for estimating a camera's position an orientation for a general class of mobile context-aware applications in architecture, engineering, an construction (AEC). Metho By stuying two classic types of natural marker-base registration algorithms, homographyfrom-etection an homography-from-tracking, an overcoming their specific limitations of jitter an rift, our metho applies two global constraints (geometric an appearance) to prevent tracking errors from propagating between consecutive frames. Results & Discussion The propose metho is able to achieve an increase in both stability an accuracy, while being fast enough for real-time applications. Experiments on both synthesize an real worl test cases emonstrate that our metho is superior to existing state-of-the-art registration algorithms. The paper also explores several AEC-applications of our metho in context-aware computing an esktop augmente reality. Keywors: information technology, AR registration, tracking, context-aware computing INTRODUCTION The ability to recover a user s pose (i.e., position an orientation within a certain coorinate system) is critical in many engineering omains such as Augmente Reality (AR), robotics, context-aware computing, an computer vision, aresse with ifferent terminology. For simplicity, we will refer to all of these as the registration problem in this paper. Despite the rapi evelopment of sensor technology such as the global positioning system (GPS) an inertial measurement units (IMU), as well as angle sensors like the igital magnetic sensor, gyroscope, an accelerometer this problem remains a challenge. A GPS signal is harly available inoors, IMU suffers from the rifting effect 1, an a magnetic sensor can be hugely affecte by the changing environment, especially in challenging environments such as construction sites with all kins of machines moving aroun, neeless to mention the sensor s annoying jitter effect. To overcome these technical insufficiencies, especially for inoor environments, infrastructurebase technology has been stuie, such as RFIDbase inoor tracking an wireless local area network (WLAN)-base inoor positioning. Yet these technologies are either costly or sensitive to the environment, an lots of work has to be put into the system calibration stage. Further, these technologies general inability to recover a user s orientation is troublesome for 3D visualization. However, beyon all of those technologies, it is very interesting to note that a human being can figure out where s/he is to a certain egree of accuracy, given that s/he is familiar enough with that specific region of environment, mostly with the help of visual clues. Following this intuition, this paper proposes a new visual registration metho calle KEG planar object tracker, which essentially recovers the pose of the user, i.e. camera, in real-time from a set of planar markers whose own poses are known, capturing the iea that our tracker is familiar enough with the environment so as to perform an estimation of position an orientation, just as humans o. Firstly, the state-of-the-art methos in visual registration will be briefly introuce. Then among these methos, two important types of planar marker-base algorithms are iscusse. The section following that explains the main contribution of this paper an efficient improvement base on the previous two classes of algorithm frameworks, leaing to the KEG tracker. Afterwars, several experiments are shown to emonstrate the superiority, uner ifferent objective quality measures, of KEG tracker. Also, two novel AEC applications that eploy KEG tracking algorithm is introuce. REVIEW OF PRIOR RELATED WORK The visual registration problem is actively stuie in the computer vision community, an several algorithms have been propose to aress it. Base on their ifferent assumptions on the environment (i.e. the surrouning worl where visual registration is going to be performe), these algorithms can be classifie into two groups 2 : known environment vs. unknown environment. The first group of algorithms is less computation-consuming an easier to esign since the only unknown is the user s pose. Because they have been well stuie, an many relate powerful algorithms have been propose in the last

2 Fig. 1. Homography-from-etection algorithm framework. two ecaes, it s more realistic to apply them for solving real-worl engineering issues. Within this class of methos, they can be further categorize into two groups: planar environment vs. non-planar environment. The first group is again easier to esign because of the simple assumption mae regaring the environment a planar structure with known visual features. An the secon group is more often applie in a controlle environment with limite space, such as a small manufacturing workspace. The authors choose to take avantage of planebase methos since planar structures are abunant in builings, construction sites, an other humanmae environments where engineering operations are conucte, which makes this type of metho very convenient to apply. In aition, a planar structure can simply be an image printe out on a piece of paper an attache to a wall/floor of a corrior, with nearly zero cost. All of those avantages make this metho ieal for application, an merit its investigation. Plane-base methos can be further classifie base on ifferent visual features they aopt: fiucial marker vs. natural marker. A fiucial marker is compose of a set of visual features that are easy to extract an provie reliable, easy to exploit measurements for the pose estimation 2. Usually those features are a set of black an white patterns forming simple geometry by circles, straight lines, or sharp corners an eges. Well known fiucial markers inclue ARToolKit 3 an the newly propose AprilTag 4. Distinct from a fiucial marker, a natural marker oes not require special preefine visual features. Instea, it treats any visual features in the same way. In this sense, almost any common image, ranging from a natural view to a company logo, can immeiately be use as a natural marker. This major ifference makes it much easier an more natural to set up a natural marker than a fiucial one. Users o not nee to separately esign special markers; they can simply take avantage of any meaningful pictures relate to the problem of interest. In aition, one major ownsie to a fiucial marker lies in the fact that it usually epens on the four corner points or eges of the marker s quarangle to o further registration estimation, which is the reason that fiucial marker-base methos will fail even if one corner is not within view. This isavantage oes not exist in natural marker-base methos; in fact natural markers o not even require a marker image to be rectangular. Again, by the funamental ifference in the way they treat input images, natural marker-base methos form two groups: one group treats each input image inepenently, which is referre to as a etectionbase metho, such as 5,6 ; the other group nees two or more consecutive images as input, which is referre to as a tracking-base metho, such as 7,8. Since our propose metho evolves from both these algorithm groups, in the following sections we will explain in etail the process framework of each type of algorithms, as well as how it inspires an is jointly aapte to our propose algorithm framework. HOMOGRAPHY FROM DETECTION In either fiucial marker-base or natural markerbase algorithms, the funamental task is to fin the transformation between the marker image plane an the current camera plane which contains that current image frame. Such a transformation, calle as homography, maps points on the marker image to their corresponing points on the current image frame with the following equation: where H is a 3x3 matrix representing the homography, ( x, y ) an ( x, y ) are the corresponing points on the two images, an s is an unknown scaling parameter. In fact, the general iea behin a plane-base registration algorithm is the fact that homography between two planes encoes the orientation an position information of one plane relative to another. From this perspective, registration is equivalent to fining the homography between the marker plane an the current camera plane. From projective geometry, one knows that with at least four point corresponences between two planes, their homography can be uniquely etermine by solving a set of linear equations 9. More complicate than fiucial marker-base algorithms, which take avantage of simple patterns to fin corresponences an then estimate homography, natural marker-base algorithms require a lot more effort to solve a corresponence problem. Fig. 1 shows the generic algorithm framework of the homography-from-etection type of methos. The gray components nee be loae or calculate once

3 Fig. 2. Homography-from-tracking algorithm framework. uring a computation, while the white components nee to be upate for each new frame. H is the homography between the current frame an the marker image. K is the camera calibration matrix storing the focal length an some other intrinsic parameters, which can be calibrate in avance. R is the rotation matrix representing camera orientation, an T is the translation vector representing the position of camera center. For each incoming image frame, the first step is to etect a set of keypoints. Also, at the very beginning, a fixe set of keypoints has to be etecte on the marker image. Interest point etection algorithms are usually applie in this step. The secon step involves a matching problem, i.e. fining corresponing points between two sets of keypoints base on their local appearance. Among the state-of-the-art algorithms, the Scale Invariant Feature Transform (SIFT) algorithm 5 is perhaps the most famous an wiely use nowaays. Although the SIFT algorithm works very well uner large variation of visual conitions, it is computationconsuming, which makes it impractical to be applie irectly in real-time applications, such as a registration problem, even after applying lot of approximation to SIFT. FERNs 6, iffers from SIFT by the requirement of an offline training stage. Only after a long perio of training using the marker image can FERNs recognize ifferent keypoints on that particular marker uner ifferent visual conitions. Although FERNs an other similar methos enjoy the high-spee avantage, their relatively low recognition rate make them less ieal in registration problem, as to be shown by our experiments. As mentione, since most of these matching algorithms exploit a local feature escriptor, i.e. using image intensity information to escribe a keypoint within only a limite neighboring region centere at that keypoint, mismatch is inevitable. In orer to avoi most of these false matches, a robust estimation algorithm, such as the famous RANom Sample An Consensus 10 (RANSAC) is usually employe to estimate the homography. Once homography is foun through matrix ecomposition techniques the camera position, the translation vector T, an orientation the rotation matrix R, can be calculate. In our algorithm, a simple yet effective metho 11 was aopte. HOMOGRAPHY FROM TRACKING As shown in Fig. 2 homography-from-tracking type of methos explore relations between consecutive frames. Since images of two such frames usually look very similar, corresponences neee for homography estimation can easily be maintaine by tracking each keypoint aroun its local neighborhoo. Thus this type of methos can circumvent the harest matching problem, since in this framework, matching between the marker keypoint an the current keypoint to get corresponences is only neee at the very beginning; after that, keypoint corresponences are maintaine by a tracking algorithm. In fact, there are two such tracking algorithms that play crucial roles in our propose metho: the Kanae-Lucas-Tomasi (KLT) feature tracker 7 an Efficient Secon-orer Minimization (ESM) algorithm. The KLT tracker s ultimate goal is to fin the feature point isplacement within two consecutive frames. It assumes that uring these two frames, the local appearance of the feature point x oes not change, an that the isplacement is small. Then in orer to fin the optimal isplacement, the algorithm formulates a least square problem to achieve a fast solution. While KLT s motion moel being fairly simple, ESM 8 uses the 2D homography as motion moel, an uses secon-orer approximation of the image function, thus leaing to a global refinement algorithm with a faster convergence rate. GLOBAL GEOMETRIC/APPEARANCE CONSTRAINTS The avantage of homography-from-etection methos lies in the fact that since they treat each image frame separately, as we have shown in Fig. 1, estimation can be totally wrong at one particular frame, an the following frames won t be affecte at all. However, the problem with methos such as SURF an FERNs is that, in orer to spee up the time-consuming matching step in etection, lots of approximations are aapte. This makes the homography estimation very unstable, resulting in a very annoying jitter effect if aopte in AR applications, i.e. the augmente object appears to be shaking in the scene. In our experiments, even if the camera is fixe, the estimate camera position an orientation coul have very large variance.

4 Fig. 3. KEG algorithm framework. In a ifferent approach, homography-from-tracking methos, such as ESM an KLT, compare the current frame with the previous one in orer to track the change. Although they benefit from the relatively higher tracking spee that improves their frame rate, one critical problem of this group of methos is that every tracker suffers from the rifting effect, i.e. the upate position of the tracke point actually iffers to some extent from its true new position, an will thus eventually fail. The rifting effect will lea to large error in homography estimation, since these rifting errors usually o not follow Gaussian istribution an shall be seen as systematic errors that are changing ynamically. Also, the greater the number of points failing to be tracke, the larger the variance that the estimate camera pose coul have. Thus the augmente objects coul be in a wrong position an shaking at the same time. Our new framework (Fig. 3) integrates the homography-from-etection an homography-fromtracking frameworks, utilizing their strong points an circumventing their short-comings. In general, our framework starts by the original homography-frometection framework. Once the marker image is etecte along with a rough estimation of the homography, we immeiately move into a coarse-tofine framework. Only when the track is somehow lost will this proceure be repeate. Within our new framework, an firstly via the original tracking algorithm (KLT), a coarse homography coul be foun. Then, it woul be by a global optimization algorithm (ESM). Finally the homography woul be use to correct the positions of the set of points to be tracke in the next frame, which is inspire by our following analysis of the cause of the rifting effect. Drifting Effect Analysis After analyzing the rifting effect in homographyfrom-tracking methos, we foun that it is actually an error accumulation issue. During the tracking between every two consecutive frames, the error introuce by any tracking algorithm is accumulate. After a while, this accumulation coul irectly lea to the tracking of a wrong local target or even to the failure of the tracker. After realizing this, one pertinent question to ask is: is there any way to correct the error before the next tracking is actually performe? It was foun that this is possible, which le to the esign of our propose framework. To gain more unerstaning of the cause of the rifting effect, the etaile error analysis is shown, as follows. Firstly, the error source of any tracking algorithm, such as KLT, can be seen as compose by the following terms: (1) where x new is the upate position estimate by KLT, x ol is the true original position, Δx xnew x ol is the true isplacement, is the systematic rifting error, an g is the rest of the error, which is assume to follow some Gaussian istribution. Here, Δx is cause by physical movement between the camera an the scene, g is mainly cause by camera CCD sensor noise, an is usually cause by the tracking algorithm an other complicate reasons, such as the fact that KLT will be affecte a lot when the camera is moving too fast, which leas to motion blur an thus violates KLT s unerlying assumptions. The secon step of homography-from-tracking methos applies RANSAC to estimate H coarse from the array of tracke points { x an their corresponing points on marker image. However, even though RANSAC can eliminate a lot of outliers if the absolute value of error g excees some threshol, an further, can eliminate the Gaussian error g by a final least-square estimation on the outlier-free subset of corresponent points, there still remains a part of systematic rifting error not hanle an thus propagate into H coarse. In the homography-from-tracking framework, neither nor g are correcte uring the upate step, so these errors are accumulate, which can cause a large rift even after a few frames of tracking. Error Correction by Global Constraints One natural way to reuce the effect of is to apply the global appearance constraint, as shown in Fig. 3, which essentially means that the original

5 Algorithm 1: KEG algorithm framework. 1. Detect N keypoints { x ref } on marker image T. 2. Apply Fiucial Marker metho (AprilTag) or Homography-from-etection algorithm (SIFT), try to fin the marker image an its corresponing homography H. If foun, go to step 6; otherwise, re-o step Take a new incoming frame I new, the last frame I ol, an the ol keypoints positions { x ol }, perform local tracking (KLT) an output new keypoints positions { x. 4. Perform robust estimation (RANSAC) on corresponent keypoint array { x ref } an { x, then output H coarse. 5. Apply global appearance constraint by ESM an output H. 6. Valiate H by similarity (zero-mean normalize cross-correlation, equation (3)) threshol. If vali, H can be output for homography ecomposition; otherwise, meaning loss-of-track, go to step Upate positions of keypoints { x by equation (2). 8. Replace I ol with I new. Replace { x ol } with { x. Go to step 3. marker image shoul look the same as the image rectifie from the current frame by estimate homography H. Before this constraint, all of the information the KLT tracker use is local, while is systematic, therefore a global optimization base on the whole marker image will theoretically eliminate all the systematic error an H coarse can serve as a goo optimization starting point. After the rifting error is eliminate when estimating the homography H, we can easily correct tracking errors an upate keypoint positions to be fille into the next tracking iteration by the homography mapping: (2) where x ref is a keypoint s position on the original marker image; we refer to this as applying the global geometric constraint (for it relies on the prior knowlege that all keypoints lie in the same plane). Since the estimate H is alreay theoretically error-free, upating using the above equation (2) instea of equation (1) prevents tracking error from propagating into the tracking of the next frame, an thus increases the tracking stability. Besies the improvement in accuracy, our algorithm also enjoys an increase in tracking spee. Because we have a global refinement step, we o not require the local tracking algorithm such as KLT to be very accurate by reucing the number of iterations of KLT algorithms that result in larger error g. Since the irect result of KLT is just a coarse homography serving as an ESM optimization starting point, a certain amount of error can be tolerate an will be theoretically eliminate after global refinement (ESM). Similarly, since the time complexity of a local tracking algorithm such as KLT is usually positively correlate to the number of points to be tracke, we can ecrease the number of keypoints to be tracke. We refer to our metho as the KEG (KLT Enhance by Global constraints) tracker, an the complete algorithm framework is escribe in Algorithm 1. It s worth noting that KLT, ESM, an RANSAC, as well as the initial etection metho (AprilTag/SURF), are replaceable components in our approach. This makes our metho very flexible an easy to be extene by new algorithms (as long as they serve the same purpose). We will offer etaile comparisons in next section showing that, even though our framework involves more steps, its performance in accuracy, stability, an spee is increase as compare to the state-of-the-art algorithms. EXPERIMENTAL RESULTS In orer to valiate our metho an compare it to state-of-the-art algorithms, we i several experiments on both real-worl an synthesize vieo sequences (in which we knew the groun-truth of the camera pose). Experiments were all conucte on a esktop computer with an eight-core 2.8 GHz Intel Core i7 CPU an similar performances were achieve on lower en computers. Also, all of the vieo sequences have a frame size of 640x480, which is the commonly aopte size of commercial webcams. In all of the test cases to be shown in the following, for the purpose of showing the necessity of the three core components in KEG local tracker (K-step), global refinement (E-step), an error correction (Gstep) an proving its superiority to state-of-the-art methos, we ran 7 ifferent algorithms over those test cases: 1. KEG with AprilTag as initialization metho (referre to as A+KEG). 2. No global appearance constraints applie; others are the same as 1 (A+K G). 3. No global geometric constraints applie; others are the same as 1 (A+KE). 4. No global constraints applie, representing homography-from-tracking metho (A+K).

6 5. AprilTag, representing fiucial marker-base metho (A). 6. AprilTag with global appearance constraints applie (A+ E). 7. FERNs, representing homography-frometection metho (FERNs). For the homography-from-etection component, we use our own C++ implementation of AprilTag, which originates from the java implementation by April Lab at the University of Michigan 4. For comparison with state-of-the-art homography-from-etection methos, we aopte the well-known an wiely use Opensource Computer Vision (OpenCV) library implementation of the FERNs metho. For the ESM metho, we use the binary library provie by INRIA at We also looke at ifferent performance metrics so as to have a comprehensive unerstaning of the performance of these algorithms: Duration: The time to process each frame, reflecting the spee of the algorithm. This metric is crucial for real-time applications. NCC: The zero-mean normalize cross-correlation between the marker image I 1 an the rectifie image I 2 by H, which can be calculate by: where n is the total number of pixels of image I 1 or I 2, an m i an i are the mean value an stanar eviation of intensity of image I 1. Obviously, if I 1 an I 2 are exactly the same, their NCC inex shoul be one, an the larger their ifference, the smaller the NCC inex. This means NCC is a goo similarity inex. This inex is also use in KEG to etermine whether it loses track or not by a simple threshol of 0.5; if at any frame the NCC inex is smaller than 0.5, it is regare as a loss-of-track frame. LOT: The total number of loss-of-track frames. This metric represents a registration algorithm s stability to some extent. T-RMS/R-RMS: The root mean square error between estimate camera position/orientation an groun truth. This metric represents the absolute accuracy of a registration algorithm, an is calculate by: (3) where is the total number of frames of a test case, T i an T ˆi are the i-th frame s groun truth an estimate position vector of imension 3x1, an E i an E ˆi are the i-th frame s groun truth an estimate Euler angle vector of imension 3x1, respectively. Since these two inices require groun truth ata, they are only examine for synthesize test cases. UOT: We also propose this new inex for estimating the extent of jitter effect of a registration algorithm, i.e. the unsmoothness-of-tracking, taking avantage of the NCC inex by where NCC i is the NCC inex of the i-th frame, which essentially means that UOT inex is the stanar eviation of the ifference between consecutive NCC inices. In MATLAB, this can be simply calculate by st(iff(ncc)). A stable registration algorithm shoul give a UOT inex value as small as possible. In all of the test cases, our marker image is compose of a 16 bits AprilTag of ID equal to zero an a natural image, the logo of University of Michigan that is rich in features. The synthesize test case is renere in OpenGL, using our marker image an a static real-worl image as backgroun. An the real-worl test case is recore using a Logitech webcam. Experiments shows that the average uration per frame of A+KEG is about 40 millisecons, which is even faster than AprilTag (60~70 millisecons). While other algorithms have very unstable processing time an are mostly slower than A+KEG, the only exception is A+KG, which is as expecte since it has no global refinement step. These results prove that the KEG metho oes fit for real-time applications with process spee of 20 frames-persecon or faster. Fig. 4 (a) shows that in most of the time, the NCC curve of A+KEG is the upper boun for the other algorithms, especially performing better than the state-of-the-art algorithm, FERNs. This means KEG tracker has the highest quality of tracking. Fig. 4 (b) an (c) emonstrate that the A+KEG metho has fewer loss-of-track frames, showing its ability to track longer, an lower UOT inex, showing its smoothness in tracking an important feature if applie in AR. Also A+KEG is more accurate, by giving less T-RMS/R-RMS errors in Fig. 4 () an (e). Notice that here we assume the raius of our synthesize marker is 20 cm (which was the real size when it was printe out on an A4 sheet of paper in the real-worl test cases). In this configuration, the KEG tracker s maximum working istance can be as far as 3 meters, an its maximum working Euler angle can be about 85 egree offset from the marker image s normal irection. If an even larger working istance is esirable, a higher resolution camera an bigger marker can be aopte.

7 (a) Real-worl case s NCC curve. (b) Real-worl case s UOT inex. (c) Real-worl case s LOT inex. () Synthesize case s T-RMS inex. (e) Synthesize case s R-RMS inex. Fig. 4. Experiment results of comparison between ifferent algorithms. APPLICATIONS As note in the Introuction, this algorithm has several potential applications in many ifferent areas. One example application we implemente is contextaware computing. Inoor context-aware computing has been stuie in AEC for its ability to spee up information elivery in many aspects, incluing construction site inspection/monitoring an facility management 1,13. Prior approaches for inoor ubiquitous tracking utilize an inertial measurement evice, which suffers from its rifting effect. By Akula et al. 1, a context-aware computing system integrate with both GPS an inertial measurement evice is evelope, requiring human intelligence to recognize certain preefine locations to manually correct the rifting error cause by the IMU. In our application, manual error correction was naturally replace by automate correction using the A+KEG metho, as shown in Fig. 5 left. The green text shows that the algorithm successfully recognizes ifferent locations by composing a natural photo (UM logo in this case) with ifferent AprilTags. Once the inspector is within the effective range of the KEG marker image, the marker is automatically etecte an then the inspector s pose relative to the marker is continuously estimate. Similar to Akula et al. s metho, both the location an orientation of these preefine markers in the global coorinate system are known an store in a atabase. Therefore the inspector s pose in the global coorinate system can be etermine, as well. Benefiting from the KEG tracker, this application can provie automatic regional rifting error correction instea of manual point correction. This application can thus further facilitate information elivery on construction sites or in inoor builing environments. Another interesting application lies in tabletop augmente reality. Fig. 5 right shows a esktop environment AR showcase of a 3D builing esign. Since the KEG tracker has the ability to quickly etect an accurately maintain the tracking of a marker image without requiring that the marker image be fully in sight (as require by ARToolkit), it can easily be aapte into tabletop collaborative AR applications 12 to support better interactive esign emonstration or visual simulation for construction planning. Note that our KEG algorithm is opensource, so the C++ coes an emonstration vieos for the above two applications coul be foun at our lab s webpage (

8 Fig. 5. A+KEG in context-aware computing (left) an table taletop AR (right). CONCLUSIONS After stuying the two ifferent types of natural marker-base registration algorithms an analyzing the cause of the rifting an jitter effects in both homography-from-tracking an homography-frometection methos, a new natural marker-base registration algorithm, KEG tracker, is propose, combining the avantages of those two, an circumventing their shortcomings. In theoretical analysis, we foun that the rifting effect is an error that occurs because of an accumulation problem. We solve the problem by applying two global constraints: a geometric an appearance constraint. Experiments on both synthesize an real-worl cases prove that the KEG metho is fast enough for real-time applications, an also more accurate an stable than state-of-the-art algorithms such as FERNs an AprilTag, even when the marker image is partially occlue or viewe from very long istance. We also explore potential applications of the new tracker, such as context-aware computing, for replacing manually rifting error correction, an augmente reality for tabletop 3D visual simulation. In the future, one irection for further research is applying more object recognition techniques so that, without the nee of composing a fiucial marker (AprilTag), our metho coul more naturally support multiple marker recognition an tracking. Extening our metho to a 3D environment, such that no planar structure assumption is neee, coul also be a very interesting research irection. In aition, specific AEC applications of the tracker can be explore, such as pose estimation for construction equipment. References 1. Akula, M., Dong, S., Kamat, V., Ojea, L., Borrell, A., an Borenstein, J., "Integration of infrastructure base positioning systems an inertial navigation for ubiquitous context-aware engineering applications," Avance Engineering Informatics, Lepetit, V., an Fua, P., "Monocular Moel-Base 3D Tracking of Rigi Objects," Founations an Trens in Computer Graphics an Vision, Kato, H., an Billinghurst, M., "Marker tracking an hm calibration for a vieo-base augmente reality conferencing system," in Proceeings of 2n IEEE an ACM International Workshop on Augmente Reality, Olson, E., "AprilTag: A robust an flexible visual fiucial system," in Proceeings of the International Conference on Robotics an Automation, Lowe, D., "Distinctive image features from scaleinvariant keypoints," International journal of computer vision, Vol. 60(2), pp , Ozuysal, M., Fua, P., an Lepetit, V., "Fast keypoint recognition in ten lines of coe," in Proceeings of IEEE Conference on Computer Vision an Pattern Recognition, Lucas, B., an Kanae, T., "An Iterative Image Registration Technique with an Application to Stereo Vision," in Proceeings of the 7th International Joint Conference on Artificial Intelligence, Benhimane, S., an Malis, E., "Real-time imagebase tracking of planes using efficient seconorer minimization," in Proceeings of International Conference on Intelligent Robots an Systems, Hartley, R., an Zisserman, A., Multiple view geometry in computer vision, Cambrige University Press, Fischler, M., an Bolles, R., "Ranom sample consensus: a paraigm for moel fitting with applications to image analysis an automate cartography," Communications of the ACM, Vol. 24(6), pp , Simon, G., Fitzgibbon, A., an Zisserman, A., "Markerless tracking using planar structures in the scene," in Proceeings of IEEE an ACM International Symposium on Augmente Reality, Dong, S., an Kamat, V., "Collaborative Visualization of Simulate Processes Using Tabletop Fiucial Augmente Reality," in Proceeings of Winter Simulation Conference, Behzaan, A., Aziz, Z., Anumba, C., an Kamat, V., "Ubiquitous location tracking for context-specific information elivery on construction sites," Automation in Construction, Vol. 17(6), pp , 2008.