======================================================================== README for Augmented Reality Sandbox (SARndbox) version 1.5 Copyright (c) 2012-2013 Oliver Kreylos ======================================================================== Overview ======== The Augmented Reality Sandbox is an augmented reality application scanning a sand surface using a Kinect 3D camera, and projecting a real- time updated topography map with topographic contour lines, hillshading, and an optional real-time water flow simulation back onto the sand surface using a calibrated projector. Requirements ============ The Augmented Reality Sandbox requires Vrui version 3.0 build 001 or newer, and the Kinect 3D Video Capture Project version 2.7 or newer. Installation Guide ================== It is recommended to download or move the source packages for Vrui, the Kinect 3D Video Capture Project, and the Augmented Reality Sandbox into a src directory underneath the user's home directory. Otherwise, references to ~/src in the following instructions need to be changed. It is also recommended to skip optional steps 4 and 6 in the following instructions. The Augmented Reality Sandbox does not need to be installed in order to be used; installation (to a system directory such as /usr/local) is only recommended if the Augmented Reality Sandbox will be used from multiple user accounts. 0. Install Vrui from ~/src/Vrui-- (see Vrui README file). 0.5 Install the Kinect 3D Video Capture Project from ~/src/Kinect- (see the Kinect 3D Video Capture Project README file). 1. Change into ~/src directory and unpack the Augmented Reality Sandbox tarball: > cd ~/src > tar xfz /SARndbox-.tar.gz - or - > tar xf /SARndbox-.tar 2. Change into the Augmented Reality Sandbox's base directory: > cd SARndbox- 3. If the Vrui version installed in step 0 was not 3.0, or Vrui's installation directory was changed from the default of ~/Vrui-3.0, adapt the makefile using a text editor. Change the value of VRUI_MAKEDIR close to the beginning of the file as follows: VRUI_MAKEDIR := /share/make Where is the installation directory chosen in step 0. Use $(HOME) to refer to the user's home directory instead of ~. 4. Optional: Adapt makefile if the Augmented Reality Sandbox is to be installed in a different location, for example /usr/local. Set INSTALLDIR to the desired target location. The Augmented Reality Sandbox will then be installed in /bin, and its configuration files will be installed in /etc (where is the value of INSTALLDIR set in the makefile). 5. Build the Augmented Reality Sandbox: > make 6. Optional: Install the Augmented Reality Sandbox in the selected target location. This is only necessary if the INSTALLDIR variable in the makefile was changed. By default, the Augmented Reality Sandbox can be run from its base directory. To install: > make install - or, if the target location is a system directory - > sudo make install This will copy all executables into /bin, and all configuration files into /etc. 7. Optional: Add directory containing the Augmented Reality Sandbox executable (~/src/SARndbox-/bin in the default installation, /bin otherwise) to the user's search path. This allows running the Augmented Reality Sandbox from any directory. Using csh or tcsh: > setenv PATH ${PATH}:~/src/SARndbox-/bin - or - > setenv PATH ${PATH}:/bin where is the target location set in the makefile. Using bash: > export PATH=${PATH}:~/src/SARndbox-/bin - or - > export PATH=${PATH}:/bin These lines can also be added to the user's .cshrc or .bashrc files to make the additions persist between logins. Use === The Augmented Reality Sandbox package contains the sandbox application itself, SARndbox, and a calibration utility to interactively measure a transformation between the Kinect camera scanning the sandbox surface, and the projector projecting onto it. The setup procedure described below also uses several utilities from the Kinect 3D video capture project. Setup and Calibration --------------------- Before the Augmented Reality Sandbox can be used, the hardware (physical sandbox, Kinect camera, and projector) has to be set up properly, and the various components have to be calibrated internally and with respect to each other. While the sandbox can be run in "trial mode" with very little required setup, for the full effect the following steps have to be performed in order: 1. (Optional) Calculate per-pixel depth correction coefficients for the Kinect camera. 2. (Optional) Internally calibrate the Kinect camera. 3. Mount the Kinect camera above the sandbox so that it is looking straight down, and can see the entire sand surface. Use RawKinectViewer from the Kinect 3D video capture project to line up the depth camera while ignoring the color camera. 4. Measure the base plane equation of the sand surface relative to the Kinect camera's internal coordinate system using RawKinectViewer's plane extraction tool. (See "Using Vrui Applications" in the Vrui HTML documentation on how to use RawKinectViewer, and particularly on how to create / destroy tools.) 5. Measure the extents of the sand surface relative to the Kinect camera's internal coordinate system using KinectViewer and a 3D measurement tool. 6. Mount the projector above the sand surface so that it projects its image perpendicularly onto the flattened sand surface, and so that the projector's field-of-projection and the Kinect camera's field-of- view overlap as much as possible. Focus the projector to the flattened average-height sand surface. 7. Calculate a calibration matrix from the Kinect camera's camera space to projector space using the CalibrateProjector utility and a circular calibration target (a CD with a fitting white paper disk glued to one surface). 8. Test the setup by running the Augmented Reality Sandbox application. Step 1: Per-pixel depth correction ---------------------------------- Kinect cameras have non-linear distortions in their depth measurements due to uncorrected lens distortions in the depth camera. The Kinect 3D video capture project has a calibration tool to gather per-pixel correction factors to "straighten out" the depth image. To calculate depth correction coefficients, start the RawKinectViewer utility and create a "Calibrate Depth Lens" tool. (See "Using Vrui Applications" in the Vrui HTML documentation on how to create tools.) Then find a completely flat surface, and point the Kinect camera perpendicularly at that surface from a variety of distances. Ensure that the depth camera only sees the flat surface and no other objects, and that there are no holes in the depth images. Then capture one depth correction tie point for each distance between the Kinect camera and the flat surface: 1. Line up the Kinect camera. 2. Capture an average depth frame by selecting the "Average Frames" main menu item, and wait until a static depth frame is displayed. 3. Create a tie point by pressing the first button bound to the "Calibrate Depth Lens" tool. 4. De-select the "Average Frames" main menu item, and repeat from step 1 until the surface has been captured from sufficiently many distances. After all tie points have been collected: 5. Press the second button bound to the "Calibrate Depth Lens" tool to calculate the per-pixel depth correction factors based on the collected tie points. This will write a depth correction file to the Kinect 3D video capture project's configuration directory, and print a status message to the terminal. Step 2: Internally calibrate the Kinect camera ---------------------------------------------- Individual Kinect cameras have slightly different internal layouts and slightly different optical properties, meaning that their internal calibrations, i.e., the projection matrices defining how to project depth images back out into 3D space, and how to project color images onto those reprojected depth images, differ individually as well. While all Kinects are factory-calibrated and contain the necessary calibration data in their firmware, the format of those data is proprietary and cannot be read by the Kinect 3D video capture project software, meaning that each Kinect camera has to be calibrated internally before it can be used. In practice, the differences are small, and a Kinect camera can be used without internal calibration by assigning default calibration values, but it is strongly recommended to perform calibration on each device individually. The internal calibration procedure requires a semi-transparent calibration target; precisely, a checkerboard with alternating clear and opaque tiles. Such a target can be constructed by gluing a large sheet of paper to a clear glass plate, drawing or ideally printing a checkerboard onto it, and cutting out all "odd" tiles using large rulers and a sharp knife. It is important that the tiles are lined up precisely and have precise sizes, and that the clear tiles are completely clean without any dust, specks, or fingerprints. Calibration targets can have a range of sizes and numbers of tiles, but we found the ideal target to contain 7x5 tiles of 3.5"x3.5" each. Given an appropriate calibration target, the calibration process is performed using RawKinectViewer and its "Draw Grids" tool. The procedure is to show the calibration target to the Kinect camera from a variety of angles and distances, and to capture a calibration tie point for each viewpoint by fitting a grid to the target's images in the depth and color streams interactively. The detailed procedure is: 1. Aim Kinect camera at calibration target from a certain position and angle. It is important to include several views where the calibration target is seen at an angle. 2. Capture an average depth frame by selecting the "Average Frames" main menu item, and wait until a static depth frame is displayed. 3. Drag the virtual grids displayed in the depth and color frames using the "Draw Grid" tool's first button until the virtual grids exactly match the calibration target. Matching the target in the depth frame is relatively tricky due to the inherent fuzziness of the Kinect's depth camera. Doing this properly will probably take some practice. The important idea is to get a "best fit" between the calibration target and the grid. For the particular purpose of the Augmented Reality Sandbox, the color frame grids can be completely ignored because only the depth camera is used; however, since calibration files are shared between all uses of the Kinect 3D video capture project, it is best to perform a full, depth and color, calibration. 4. Press the "Draw Grid" tool's second button to store the just-created calibration tie point. 5. Deselect the "Average Frames" main menu entry, and repeat from step 1 until a sufficient number of calibration tie points have been captured. The set of all tie points already selected can be displayed by pressing the "Draw Grid" tool's third button. After all tie points have been collected: 6. Press the "Draw Grid" tool's fourth button to calculate the Kinect camera's internal calibration parameters. These will be written to an intrinsic parameter file in the Kinect 3D video capture project's configuration directory. Step 3: Mount the Kinect camera above the sandbox ------------------------------------------------- In theory, the Kinect camera can be aimed at the sand surface from any position and/or angle, but for best results, we recommend to position the camera such that it looks straight down onto the surface, and such that the depth camera's field-of-view exactly matches the extents of the sandbox. RawKinectViewer can be used to get real-time visual feedback while aligning the Kinect camera. Step 4: Measure the base plane equation of the sand surface ----------------------------------------------------------- Because the Kinect camera can be aimed at the sand surface arbitrarily, the Augmented Reality Sandbox needs to know the equation of the "base plane" corresponding to the average flattened sand surface, and the "up direction" defining elevation above or below that base plane. The base plane can be measured using RawKinectViewer and the "Extract Planes" tool. Flatten and average the sand surface such that it is exactly horizontal, or place a flat board above the sand surface. Then capture an average depth frame by selecting the "Average Frames" main menu entry, and wait until the depth image stabilizes. Now use the "Extract Planes" tool to draw a rectangle in the depth frame that *only* contains the flattened sand surface. After releasing the "Extract Planes" tool's button, the tool will calculate the equation of the plane best fitting the selected depth pixels, and print two versions of that plane equation to the terminal: the equation in depth image space, and the equation in camera space. Of these, only the second is important. The tool prints the camera-space plane equation in the form x * (normal_x, normal_y, normal_z) = offset This equation has to be entered into the sandbox layout file, which is by default called BoxLayout.txt and contained in the Augmented Reality Sandbox's configuration directory. The format of this file is simple: the first line contains the sandbox's base plane equation in the form (normal_x, normal_y, normal_z), offset The plane equation printed by the "Extract Planes" tool only needs to be modified slightly when pasting it into the sandbox layout file: the "x *" part has to be removed, and the equal sign has to be replaced by a comma. The other four lines in the sandbox layout file are filled in in the next calibration step. The base plane equation defines the zero elevation level of the sand surface. Since standard color maps equate zero elevation with sea level, and due to practical reasons, the base plane is often measured above the flattened average sand surface, it might be desirable to lower the zero elevation level. This can be done easily be editing the sandbox layout file. The zero elevation level can be shifted upwards by increasing the offset value (the fourth component) of the plane equation, and can be shifted downwards by decreasing the offset value. The offset value is measured in cm; therefore, adding 10 to the original offset value will move sea level 10 cm upwards. Step 5: Measure the extents of the sand surface ----------------------------------------------- The Augmented Reality Sandbox needs to know the lateral extents of the visible sand surface with respect to the base plane. These are defined by measuring the 3D positions of the four corners of the flattened average sand surface using KinectViewer and a 3D measurement tool, and then entering those positions into the sandbox layout file. Start KinectViewer, and create a 3D measurement tool by assigning a "Mouse -> Screen Projector" tool to some button, and then a "Measurement Tool" tool to the same button. Then measure the 3D positions of the four corners of the flattened sand surface in the order lower left, lower right, upper left, upper right; in other words, form a mirrored Z starting in the lower left. To measure the positions correctly, change the view until the sand surface is aligned to the screen plane, i.e., until the crosshairs displayed while the viewpoint is changed exactly touch the virtual sand surface along its entire extent. Then move the mouse to the four corners of the virtual sand surface, and click the "Measurement Tool" tool's button once each time. The measurement tool will save its measurements in a file in the current directory, with a name of MeasurementToolXXXX.dat. This file will contain one line per collected measurement point, prefixed by "Navigational position" followed by the point's x, y, and z coordinates. Paste the coordinates of each point into the sandbox layout file in the order mentioned above. Afterwards, the measurement file can be deleted. Step 6: Mount the projector above the sandbox --------------------------------------------- Just like with the Kinect camera, the Augmented Reality Sandbox is capable of dealing with arbitrary projector alignments. As long as there is some overlap between the Kinect camera's field-of-view and the projector's projection area, the two can be calibrated with respect to each other. However, for several reasons, it is best to align the projector carefully such that it projects perpendicularly to the flattened average sand surface. The main reason is pixel distortion: if the projection is wildly off-axis, the size of projected pixels will change sometimes drastically along the sand surface. While the Augmented Reality Sandbox can account for overall geometric distortion, it cannot change the size of displayed pixels, and the projected image looks best if all pixels are approximately square and the same size. Some projectors, especially short-throw projectors, have off-axis projections, meaning that the image is not centered on a line coming straight out of the projection lens. In such cases, perpendicular projection does not imply that the projector is laterally centered above the sandbox; in fact, it will have to be mounted off to one side. The criterion to judge perpendicular projection is that the projected image appears as a rectangle, not a trapezoid. We strongly recommend against using any built-in keystone correction a particular projector model might provide. The Augmented Reality Sandbox corrects for keystoning internally, and projector-based keystone correction works on an already pixelated image, meaning that it severely degrages image quality. Never use keystone correction. Align the projector as perpendicularly as possible, and let the Augmented Reality Sandbox handle the rest. The second reason to aim for perpendicular projections is focus. Projector images are focused in a plane perpendicular to the projection direction, meaning that only a single line of the projected image will be in correct focus when a non-perpendicular projection is chosen. Either way, after the projector has been mounted, we recommend to focus it such that the entirety of the flattened average sand surface is as much in focus as possible. On a tangential note, we also strongly recommend to only run projectors at their native pixel resolutions. Most projector models will support a wide range of input video formats to accomodate multiple uses, but using any resolution besides the one corresponding to the projector's image generator is a very bad idea because the projector will have to rescale the input pixel grid to its native pixel grid, which causes severe degradation in image quality. Some projectors "lie" about their capabilities to seem more advanced, resulting in a suboptimal resolution when using plug&play or automatic setups. It is always a good idea to check the projector's specification for its native resolution, and ensure that the graphics card uses that resolution when the projector is connected. Step 7: Calculate the projector calibration matrix -------------------------------------------------- The most important step to create a true augmented reality display is to calibrate the Kinect camera capturing the sand surface and the projector projecting onto it with respect to each other, so that the projected colors and topographic contour lines appear exactly in the right place. Without this calibration, the Augmented Reality Sandbox is just a sandbox with some projection. This calibration step is performed using the CalibrateProjector utility, and a custom calibration target. This target has to be a flat circular disk whose exact center point is marked in some fashion. We recommend to use an old CD, glue a white paper disk of the proper size to one side, and draw two orthogonal lines through the CD's center point onto the paper disk. It is important that the two lines intersect in the exact center of the disk. The calibration procedure is to place the disk target into the Kinect camera's field-of-view in a sequence of prescribed positions, guided by the projector. When CalibrateProjector is started, it will first capture a background image of the current sand surface; it is important that the surface is not disturbed during or after this capture step, and that no other objects are between the Kinect camera and the sand surface. Afterwards, CalibrateProjector will collect a sequence of 3D tie points. For each tie point, it will display two intersecting lines. The user has to position the disk target such that the projected lines exactly intersect in the disk's center point, and such that the disk surface is parallel to the flattened average sand surface, i.e., the base plane that was collected in a previous calibration step. It is important to place the disk at a variety of elevations above and ideally below the base surface to collect a full 3D calibration matrix. If all tie points are in the same plane, the calibration procedure will fail. The exact procedure is as follows: 1. Start CalibrateProjector and wait for it to collect a background frame. Background capture is active while the screen is red. It is essential to run CalibrateProjector in full-screen mode on the projector, or the resulting calibration will be defective. See the Vrui user's manual on how to force Vrui applications to run at the proper position and size. When started, CalibrateProjector must be told the exact pixel size of the projector's image using the -s command line option. Using a wrong pixel size will result in a defective calibration. The recommended BenQ short-throw projector has 1024x768 pixels, which is also the default in the software. In other words, when using an XGA-resolution projector, the -s option is not required. 2. Create a "Capture" tool and bind it to two keys (here "1" and "2"). Press and hold "1" and move the mouse to highlight the "Capture" item in the tool selection menu that pops up. Then release "1" to select the highlighted item. This will open a dialog box prompting to press a second key; press and release "2". This will close the dialog box. Do not press "1" again when the dialog box is still open; that will cancel the tool creation process. This process binds functions to two keys: "1" will capture a tie point, and "2" will re-capture the background sand surface. "2" should only be pressed if the sand surface changes during the calibration procedure, for example if a hole is dug to capture a lower tie point. After any change to the sand surface, remove the calibration object and any other objects, press "2", and wait for the screen to turn black again. 3. Place the disk target at some random elevation above or below the flattened average sand surface such that the intersection of the projected white lines exactly coincides with the target's center point. 4. Remove your hands from the disk target and confirm that the target is seen by the Kinect camera. CalibrateProjector will display all non-background objects as yellow blobs, and the object it identified as the calibration target as a green blob. Because there is no calibration yet, the green blob corresponding to the disk target will not be aligned with the target; simply ensure that there is a green blob, that it is circular and stable, and that it matches the actual calibration target (put your hand next to it, and see if the yellow blob matching your hand appears next to the green blob). 5. Press the "Capture" tool's first button ("1"), and wait until the tie point is captured. Do not move the calibration target or hold any objects above the sand surface while a tie point is captured. 6. CalibrateProjector will move on to the next tie point position, and display a new set of white lines. Repeat from step 3 until all tie points have been captured. Once the full set has been collected, CalibrateProjector will calculate the resulting calibration matrix, print some status information, and write the matrix to a file inside the Augmented Reality Sandbox's configuration directory. The user can continue to capture more tie points to improve calibration as desired; the calibration file will be updated after every additional tie point. Simply close the application window when satisfied. Additionally, after the first round of tie points has been collected, CalibrateProjector will track the calibration target in real-time and indicate its position with red crosshairs. To check calibration quality, place the target anywhere in or above the sandbox, remove your hands, and ensure that the red crosshairs intersect in the target's center. Step 8: Run the Augmented Reality Sandbox ----------------------------------------- At this point, calibration is complete. It is now possible to run the main Augmented Reality Sandbox application, SARndbox, without any required command line options. It is, however, important to run the application in full-screen mode on the projector, or at least with the exact same window position and size as CalibrateProjector in step 7. If this is not done correctly, the calibration will not work as desired. To check the calibration, observe how the projected colors and topographic contour lines exactly match the physical features of the sand surface. If there are discrepancies between the two, repeat calibration step 7 until satisfied. On a typical 40"x30" sandbox, where the Kinect is mounted approximately 38" above the sandbox's center point, and using a perpendicularly projecting 1024x768 projector, alignment between the real sand surface and the projected features should be on the order of 1 mm. SARndbox provides a plethora of configuration files and command line options to fine-tune the operation of the Augmented Reality Sandbox as desired. Run SARndbox -h to see the full list of options and their default values, or refer to external documentation on the project's web site. Note on water simulation ------------------------ Without the real-time water simulation, the Augmented Reality Sandbox has very reasonable hardware requirements. Any current PC with any current graphics card should be able to run it smoothly. The water simulation, on the other hand, places extreme load even on high-end current hardware. We therefore recommend to turn off the water simulation, using the -ws 0.0 0 command line option to SARndbox, or reducing its resolution using the -wts command line option with small sizes, e.g., -wts 200 150, for initial testing or unless the PC running the Augmented Reality Sandbox has a top-of-the line CPU, a high-end gaming graphics card, e.g., and Nvidia GeForce 680, and the vendor-supplied proprietary drivers for that graphics card.