Hyrid Tracking System for Outdoor Augmented Reality Stelian Persa and Pieter Jonker Pattern Recognition Group, Technical University Delft Lorentzweg 1,Delft, 2628 CJ The Netherlands {stelian,pieter}@ph.tn.tudelft.nl Astract - Almost all Augmented Reality (AR) systems work indoors. Outdoor AR systems offer the potential for new application areas. The iggest single ostacle to uilding effective AR systems is the lack of accurate wide-area sensors for trackers that report the locations and orientations of ojects in an environment. Active (sensor-emitter) tracking technologies require powered-device installation, limiting their use to prepared areas that are relative free of natural or man-made interference sources. The hyrid tracker comines rate gyros and accelerometers with compass and tilt orientation sensor and GPS system. Sensor distortions, delays and drift required compensation to achieve good results. The measurements from sensors are fused together to compensate for each other's limitations. Analysis and experimental results demonstrate the system effectiveness. Keywords Outdoor navigation system, hyrid tracker system, sensor information fusion 1 INTRODUCTION The paper presents a field experiment for a low-cost strapdown-imu(inertial Measurement Unit)/GPS comination, with data processing for the determination of 2-D components of position (trajectory), velocity and heading. In the present approach we have neglected earth rotation and gravity variations, ecause of the poor gyroscope sensitivities of our low-cost ISA (Inertial Sensor Assemly) and ecause of the relatively small area of the trajectory. The scope of this experiment was to test the feasiility of an integrated GPS/IMU system of this type and to develop a field evaluation procedure for such a comination. Position and Orientation Tracking is used in Virtual Environments (VE) where the orientation and the position of a real physical oject is required. Specifying a point in 3 D requires the Cartesian coordinates x, y, and z. However, VE applications manipulate entire ojects and this requires the orientation to e specified y three angles known as pitch (elevation), roll, and yaw (azimuth). 1
With a plethora of different graphics applications that depend on motion-tracking technology for their existence, a wide range of interesting motion-tracking solutions have een invented. Surveys of magnetic, optical, acoustic, and mechanical tracking systems are availale in [1],[2]. Examples of such systems use magnetic, optical, radio, and acoustic signals. Passive-target systems use amient or naturally occurring signals. Examples include compasses sensing the Earth s field and vision systems sensing intentionally placed fiducials ( e.g., circles, squares) or natural features. Inertial systems are completely self contained, sensing physical phenomena created y linear acceleration and angular motion. Many HMD applications only require motion over a small region, and these traditional tracking approaches are usale, although there are still difficulties with interference, line-of sight, jitter and latency[5],[6]. New availale gyroscope and inertial systems represent a etter solution for the tracking system. However, a drift-corrected inertial tracking system is only ale to track a 3-DOF orientation. To correct positional drift in a 6-DOF inertial tracking system, some type of range or earing measurements to fiducial points in the environment is required. Each tracking approach has limitations. Noise, caliration error, and the gravity field impart errors on these signals, producing accumulated position and orientation drift. Position requires doule integration of linear acceleration, so the accumulation of position drift grows as the square of elapsed time. Orientation only requires a single integration of rotation rate, so the drift accumulates linearly with elapsed time. Hyrid systems attempt to compensate for the shortcomings of each technology y using multiple measurements to produce roust results. The paper is organized as follows. Section 2 descries our approach. Section 3 presents the system overview, sensor caliration and sensor fusion and filtering. The results and conclusions are presented in Section 4. 2 Approach Outdoor AR applications have rarely een attempted ecause uilding an effective outdoor AR system is much more difficult than uilding an indoor system. First, fewer resources are availale outdoors. Computation, sensors and power are limited to what a user can reasonaly carry. Second, we have little control over the environment outdoors. In an indoor system, one can carefully control the lighting conditions, select the ojects in view, add strategically located fiducials to make the tracking easier, etc. But modifying outdoor locations to that degree is unrealistic, so many existing AR tracking strategies are invalid outdoors. Finally, the range of operating conditions is greater outdoors. The amient light an outdoor display. No single tracking technology has the performance required to meet the stringent needs of outdoors AR. However, appropriately comining multiple sensors may lead to a viale solution faster than waiting for any single technology to solve the entire prolem. The system descried in this paper is our first step in this process. To simplify the prolem, we assume the real-world ojects are distant (e.g., 50+ meters), which allows the use of GPS for position tracking. Then we focus on the largest remaining sources of registration 2
error (misalignments etween virtual and real): the dynamic errors caused y lag in the system and distortion in the sensors. Compensating for those errors means stailizing the display against user motion. We do this y a hyrid tracker comining rate gyros with a compass and tilt orientation sensor. The inertial data are processed in a strapdown mechanization [3],[7], ased on the following expression for a one-component specific force in a ody reference system (see Figure 1, that explains the forces considered, acting upon the seismic mass of the accelerometer), Figure1. Specific force as a function of acceleration components along a reference system firmly attached to the moving ody (x-axis) as a function of the linear acceleration a x, the apparent centripetal acceleration a cf_x and the corresponding axial component of the static gravitational acceleration g x (the superscripts denote the vector components in the ody reference system): f m_x = ax + acf _ x gx (1) The corresponding vectorial form (with the specific force vector now denoted y a and the correction terms of centripetal and gravity acceleration expressed in the ody coordinate system) is: n a = a ω v + C g (2) with: ω = the angular velocity vector, v = the velocity vector, given in the coordinate system, and C n = the rotation matrix from the local coordinate system n to the ody coordinate system. The flow-chart of the strapdown navigation algorithm implementing the equation presented aove is presented in Figure 2. n 3
Ax Ay Az Acceleroneters Signal Correction - scale factor - ias - drift - temperature - nonorthogonality Centrifugal Force Correction wx_ vx _ wy_ vy_ w z_ vz_ Gravity Correction sθ + g cθ* sψ cθ * cψ a a a x_ y_ z_ t Acceleration Integration a( τ) dτ + 0 v v v v t x_ y_ z_ ROTATION C n vx vy vz _ n _ n _ n DGPS Position Information t Rate Integration v ( τ) dτ + n s n t 0 Gx Gy Gz Gyroscope Signal Correction - scale factor - ias - drift - temperature - nonorthogonality wx wy wz _ Attitude Integration ψ& & θ = & φ [ f ( ψ, θ, φ) ] ψ θ φ ψ θ φ Rotation Matrix C n = 2 [ f ( ψ, θ, φ) ] (C n ) T Inclinometer + Magnetometer Figure 2. Flow-chart of the strapdown mechanization We neglected the g - variations and the Earth rotation rate, ecause of the small dimensions of the test area, of the relative low car velocities (aout 1 m/s) and of the reduced rate sensitivity of the used gyroscopes. Also we neglect the small Coriolis force acting on the moving mass as a consequence of the rotation of the inertial sensors case. 3 System 3.1 Overview Figure 3 shows the system dataflow. Three sets of sensors are used: the Garmin GPS 25 LP receiver, a Precision Navigation TCM2 compass and tilt sensor, and three laser FOG (Fier Optic Gyro) rate gyroscopes (±200 degrees per second range) and three accelerometer comined in DMU-FOG sensor from Crosow. GPS 25 TCM2 DMU- FOG RS-232 I-Glass HMD 300 MHz Pentium II Laptop VGA Video Figure 3. System dataflow 4
The Garmin GPS provides outputs at 1 Hz, with 10 meters typical error, and 2-3 meter typical error in DGPS configuration. The TCM2 updates at 16 Hz and claims ±0.5 degrees of error in yaw. The gyros are analog devices which we sample at 100 Hz internally, and send via serial line. The other two sensors are also read via serial lines. An Asus 300 MHz Pentium II laptop PC reads the sensors. Section 3.2 descries the sensor distortions and caliration required. The DGPS sensor directly provides the position, ut the other two sensor outputs are fused together to determine the orientation, as descried in Section 3.3. The user location will e then passed to the renderer for display. The display is a inocular, color optical see-through HMD (I-Glass) with VGA resolution that will e rigid mounted with the sensors in order to provide a rigid relationship etween the HMD and the sensors. The software that reads from data from serial ports and fuse the data is a near real time set of threads and processes running under Windows 98. 3.2 Sensor Caliration Compass Caliration: We found the TCM2 had significant distortions in the heading output provided y the compass, requiring a sustantial caliration effort. Besides the constant magnetic declination, the compass is affected y local distortions of Earth s magnetic field. With a non-ferrous mechanical turntale is possile to measure these errors. The distortions can have peak-to-peak values of aout 2 degrees. Unfortunately, it is difficult to uild a working AR display that does not place some sources of magnetic distortion in the general vicinity of the compass. In the real system, compass errors can have peak-to-peak values of 5 degrees [8]. Fortunately TCM2 has an internal caliration procedure which can take in account a static distortion of magnetic field. For dynamic distortions the TCM2 provides us with an alarm signal, which is active when such error occurs, and then we can ignore the compass measurement and rely only on gyro. Gyroscope Caliration: We measured the ias of each gyroscope y averaging several minutes of output while the gyros were kept still. For scale, we used the specified values in the manufacturer s test sheets for each gyro. Using the caliration data for the inertial sensor assemly (ias, linear scale factors, gyroscopes triad non-orthogonality) delivered from the manufacturer and the supplementary caliration measurements made in our laoratory the error model of the inertial sensors is validated. The most important measurements are: the evaluation of the noise ehavior of the inertial data sets, static gyro calirations - to determine the supplementary non-linear terms of the static transfer characteristics, considered only to degree 2 -, as well as the estalishment of the non-linear time and temperature ehavior of the gyro s drift and scale factors and the nonorthogonality of the gyro s triad. Sensor Latency Caliration: The gyro outputs change quickly in response to user motion, and they are sampled at 100 Hz. In contrast, the TCM2 responds slowly and is read at 16 Hz over a serial line. Therefore, when TCM2 and gyro inputs are read simultaneously, there is some unknown difference in the times of the physical events they each represent. It is possile to determine the relative latency y integrating the gyro 5
outputs and compare with compass readouts y shifting one data in time till they est match. We took in account the relative latency y attaching to each sensor readout a time tag otained using Pentium II RTDS register, which operates at the processor frequency. This will e taken in account in fusion step. 3.3 Sensor fusion and filtering The goal of sensor fusion is to estimate the angular position and rotation rate of the head from the input of the TCM2 and the three gyroscopes. This position is then extrapolated one frame into the future to estimate the head orientation at the time the image is shown on the see-through display. To predict the head orientation one frame into the future, we use a linear motion model: we simply add to the current orientation the offset implied y the estimated rotational velocity. This is done y converting the orientation (the first 3 terms of x) to quaternions and using quaternion multiplication to comine them. We will incorporate more sophisticated predictors in the future. 4 Results and conclusions For moderate head rotation rates (under ~100 degrees per second) the largest registration errors we usually oserved were ~2 degrees, with average errors eing much smaller. The iggest prolem was the heading output of the compass sensor drifting with time. The output would drift y as much as 5 degrees over a few hours, requiring occasional recaliration to keep the registration errors under control. The magnetic environment also could influence the compass error, for short time we can compensate that y using only the gyro readings. In the paper some preliminary results are presented from a GPS-aided integrated trajectory solution for a low-cost strapdown mechanized IMU. The precise DGPS reference trajectory will enale the elaoration of a post-processing field evaluation methodology for the low-cost strapdown IMU. The otained results encourage to more comprehensive investigations: drift modeling of the inertial sensors in the alignment procedure, caliration of the inertial sensors error sources. Because we were primarily interested to estalish the integrated system feasiility, we have not modeled too extensively the actual inertial sensors. We intend to extend our analysis in order to achieve higher precision of the integrated solutions y the using an extended Kalman filter (EKF). Accelerometer iases, gyroscope drifts and inertial sensor scale-factor errors could e included together with appropriate stochastic models - in order to etter compensate for the systematic sensor errors. Furthermore, an increase of the inertial data acquisition rate would permit a etter approximation of the non-linear dynamic model y a linear one. Finally, for a complete dynamic model one could consider the g-variations and the 6
influence of the earth rotation, which enales the application of that analyze to more accurate IMUs, too. REFERENCES 1. Christine Younglut, Ro E. Johnson Sarah H. Nash, Ruth A. Wienclaw: Review of Virtual Environment Interface Technology, Institute for Defence Analyses, 1994 2. J. Borenstein, H.R. Everett, L. Feng. Where am I? Sensors and Methods for Moile Root Positioning, University of Michigan, 1997 3. Titterton, D., H., Weston, J., L.: Strapdown inertial navigation technology, IEE Books, Peter Peregrinus Ltd., UK, 1997 4. Jay A. Farrell, M. Barth, "The Gloal Positioning System & Inertial Navigation", McGraw-Hill, 1999 5. Foxlin, Eric, "Inertial Head-Tracker Sensor Fusion y a Complementary Separate-Bias Kalman Filter", Proceedings of VRAIS 96 (Santa Clara, CA, 30 March - 3 April 1996), 185-194 6. Foxlin, Eric, Mike Harrington, and George Pfeiffer, "Constellation: A Wide-Range Wireless Motion-Tracking System for Augmented Reality and Virtual Set Applications", Proceedings of SIGGRAPH 98 (Orlando, FL, 19-24 July1998), 371-378 7. R. Doroantu, "Field Evaluation of a Low-Cost Strapdown IMU y means GPS", Ortung und Navigation, 1/1999, DGON, Bonn 8. Azuma Ronald, Bruce Hoff, Howard Neely III, Ron Sarfaty, "A Motion-Stailized Outdoor Augmented Reality System", Proceedings of IEEE VR '99, Houston, TX, 13-17 March 1999, 252-259 7