GPS and Low Cost INS Integration for Positioning in the Urban Environment

Christopher Hide and Terry Moore Institute of Engineering Surveying and Space Geodesy, University of Nottingham

BIOGRAPHY Christopher Hide is a Research Fellow at the Institute of Engineering Surveying and Space Geodesy at the University of Nottingham. He has a degree in Mathematics and Topographic Science from the University of Wales, Swansea and a PhD in engineering surveying from the University of Nottingham. He has been working with GPS and INS integrated systems for 6 years. Terry Moore is Director of the IESSG at the University of Nottingham; where he is the Professor of Satellite Navigation. He holds a PhD degree in Space Geodesy from the University of Nottingham. He has over 20 years research experience with GNSS and is a consultant and adviser to government organizations and industry. He is a Fellow, and a Member of Council, of the Royal Institute of Navigation. ABSTRACT GPS and INS integrated systems are expected to become widely available as a result of low cost MicroElectroMechanical Sensor (MEMS) technology. However, the current performance achieved by low cost sensors is still relatively poor due to the large inertial sensor errors. This is particularly prevalent in the urban environment where there are significant periods of restricted sky view. Currently GPS and low cost INS are commonly integrated using a loosely coupled architecture which is suitable in most applications where there is good satellite availability. However, significant performance improvements can be made in the urban environment by integrating GPS at the measurement level using the tightly coupled approach. Furthermore, when it is not critical to obtain the position solution in real-time, a Kalman filter smoothing algorithm can be applied to post process the data. Kalman filter smoothing is rarely considered in the case of low cost INS, however it can result in significant performance improvements, particularly for low cost INS integration. This paper investigates the performance of a low cost Crossbow IMU integrated with GPS

measurements in the urban environment. The tightly coupled integration filter is combined with a Kalman filter smoothing algorithm to obtain a high quality integrated solution. The results obtained in this paper demonstrate that a GPS and low cost INS system is capable of meeting the performance requirements for a number of applications where GPS availability is restricted. Furthermore, by integrating GPS and INS data using a Kalman filter smoother, it is demonstrated that in post processing, high performance positioning can be achieved for a range of applications where low cost and high quality, continuous positioning is required. INTRODUCTION GPS and low cost INS integrated systems are expected to become more widespread as a result of the availability of low cost inertial Micro-ElectroMechanical Sensors (MEMS). GPS and INS integrated systems potentially provide a high data rate, high accuracy position and orientation solution that can work in all environments, particularly those where there is restricted satellite availability. Unfortunately, the performance obtained from current low cost inertial sensors is still relatively poor due to error sources such as random noise, biases and scale factor errors. Therefore these sensors are typically restricted to applications where there is good continuous satellite availability. There have been a number of recent approaches to improving the performance of GPS and low cost INS integration. For example, Nassar et al (2004) investigate the effect of improving the modelling of the IMU errors in the Kalman filter. Different integration filters have also been investigated for example, Shin and El-Sheimy (2004) and van der Merwe and Wan (2004) have investigated the Unscented Kalman Filter and Noureldin et al (2004) have considered using Neural Networks. More traditional approaches of improving GPS and INS with low cost include increasing the number of measurements used in the Kalman filter to improve the observability of errors. This can be achieved by incorporating measurements from, for example, a

magnetometer, odometer or multiple GPS antenna system. However the additional cost required for such sensors is contrary to the notion of using low cost INS. Alternatively, there exist two widely used approaches for improving integrated performance: the tightly coupled integration filter and Kalman filter smoothing. To achieve high accuracy positioning in areas where there is restricted GPS coverage, it is essential to integrate the GPS and INS sensors at the measurement level. This means that the range and range rate measurements from the GPS receiver are integrated with the INS in a tightly coupled fashion instead of using the traditional loosely coupled approach were positions and velocities are used to update the INS. This offers the significant advantage in that GPS measurements can be used to update the INS even when there are less than four satellites in view. When there are three or less satellites available, the INS position will still continue to drift, although the remaining GPS measurements can still be used to significantly reduce the error growth of the INS. In addition to using the tightly coupled approach, the integration Kalman filter can be applied in postprocessing (or potentially near real-time) using a Kalman filter smoothing algorithm. For the application of GPS and INS integration, this has a significant advantage, particularly during periods of limited GPS availability. This is as a result of using measurements from either side of the period where there are limited GPS measurements. Since the INS error grows with time since the last GPS update, the time is significantly reduced by processing the data in two directions. The smoothing algorithm also offers the advantage that all of the data is available to the filter so the inertial sensor errors are also better estimated. This paper investigates the combined use of tight coupled integration and Kalman filter smoothing to provide a high accuracy position and orientation solution using GPS and low cost INS. The urban environment is particularly challenging for GPS and INS sensors due to the infrequent measurement updates that are available. Kalman filter smoothing is often underused because of the focus of using GPS and INS in real-time applications. However, there are a significant number of applications that only require the position and orientation solution for post mission analysis. For example, surveying applications such as inventory management typically only require a solution using post processing. These applications often occur in dense urban areas where GPS signals are frequently obstructed. There is also increasing interest in the use of terrestrial georeferencing applications such as photogrammetric surveying and laser scanning (see, for example, Tayala et al, 2004). These applications also require the estimation of orientation from the INS which can be improved by post processing the data. Other applications that can operate in post

processing are vehicle performance testing such as for racing cars or product testing, and vehicle tracking applications. Since these applications are ground based, there are often significant periods of restricted satellite availability for GPS. A further application could be for road user charging where high accuracy performance is required in urban conditions and real-time operation is not essential. Therefore there already exist a number of applications that could potentially make use of the increased accuracy afforded by post-processed GPS and low cost INS measurements. Furthermore, the emergence of such a positioning sensor could potentially enable new applications that were not previously possible, primarily due to the cost of such technology. TIGHT INTEGRATION The most commonly used method of integrating GPS and INS is to use GPS position and velocity measurements to correct errors in the INS. This is generally referred to as loosely coupled integration. This is because the GPS and INS are treated as individual systems. This algorithm is also sometimes referred to as decentralised filtering since two Kalman filters are used: one to process the GPS measurements, and a second to perform the integration. This algorithm is commonly used due to its simplicity and ease of hardware implementation. However, it is clear that for applications where there is a restricted view of the sky, there is a significant improvement that can be made by combining the two systems at the measurement level. The centralised filtering algorithm uses a single Kalman filter to integrate GPS and INS measurements. This is referred to as tight integration because the INS measurements are also used to aid the GPS processing. Figure 1 describes the centralised filtering algorithm. The figure shows that the raw measurements are collected from the IMU and are converted to position, velocity and attitude measurements in the desired coordinate system using the INS mechanization algorithms. The position

Figure 1 Centralised GPS/INS filter

is then used to form predicted range measurements (corrected for the lever arm offset from the GPS antenna to the INS) to each of the satellites. These ranges are then differenced with the raw GPS measurements to estimate the INS errors using a single integration filter. Advantages of this approach include the ability to check the quality of the GPS measurements before they are used in the filter. Also, in the case when carrier phase GPS measurements are used, the INS measurements will be used to aid the ambiguity resolution algorithm. However, the principle advantage of this algorithm is that GPS measurements can still be used to update the INS when less than four satellites are available. The centralised filtering algorithm is not commonly used to integrate GPS and INS simply because of its additional complexity over the loosely coupled approach. KALMAN FILTER SMOOTHING Kalman Filter Smoothing is a post-processing algorithm that is used when all of the data is available. There are principally three types of Kalman filter smoothing known as fixed-interval smoothing, fixed-point smoothing, and fixed-lag smoothing (Gelb, 1977). Fixed-point smoothing is used to estimate a specific point in a dataset and is not considered further here. Fixed-interval smoothing estimates the states at each of the points in a data set when all the data have been collected whereas fixed-lag smoothing can be applied in near real-time. The algorithms described in this paper are obtained using fixed-interval smoothing, however the algorithm can also be modified to be used as a fixed-lag smoother for near real-time operation. Kalman filter smoothing is particularly advantageous when it is applied to GPS/INS integration during gaps in the availability of the GPS measurements. The principle of the fixed-interval smoothing algorithm during GPS outages is described in Figure 2. The figure shows that two solutions are computed in the forward and backward directions: the red line denotes the estimated error of the

forward filter estimate, and the green line denotes the estimated error of the reverse filter. The figure shows the occurrence of a data outage in the middle of the dataset. When the forward and backward position estimates are computed, the INS position error quickly increases over time which is particularly so for systems that use a low cost INS (for example, it is not uncommon to experience position errors of up to 100m during GPS outages of 60s using MEMS technology, particularly when infrequent GPS updates are available). When the Kalman filter smoothing algorithm is applied to the data, the error is significantly reduced across the data outage interval. The figure shows that Kalman filter smoothing offers a significant improvement in the estimation error over the conventional forward-only approach. Furthermore, it can be demonstrated that the smoothed solution mean square estimation error will always be less than or equal to the forward-only solution (Gelb, 1977). RAUCH-TUNG-STRIEBEL The Rauch-Tung-Striebel (RTS) algorithm provides an efficient method for implementing Kalman filter smoothing. The RTS algorithm greatly reduces the computational effort required for Kalman filter smoothing since it only requires the full Kalman filter to be implemented in the forward direction. At each measurement prediction and update step in the forward filter, the states and covariance are stored. In the case of the GPS/INS Kalman filter where the error estimates are fed back to the INS at every epoch, the smoothed estimate of the states is computed using the equation:
( xk |N = xk + ) + K k xk +1|N

(1)

(+ where xk|N is the smoothed solution at the epoch k, xk ) is

the updated state estimate from the forward filter, and K k is the Kalman filter gain. The gain is computed from the covariance of the forward filter using the equation,
K k = Pk( + ) T ( Pk(+1) ) 1

(2)

where Pk(+ ) is the updated covariance at the epoch k, and

T Pk | N = Pk( + ) + K k Pk +1| N Pk(+1) K k

(3)

However, the covariance does not have to be computed in order to compute the smoothed state, hence if not required, this equation can be ignored. Figure 2 Advantage of Kalman filter smoothing during GPS outage (adapted from Gelb, 1977)

The RTS algorithm can also potentially be used in nearreal-time by running the smoothing algorithm on short periods of data throughout during the data collection. For example, after every significant GPS outage, the algorithm could be applied once GPS availability returns. The lag at which with smoother operates does not have to be fixed which means smoothing could be applied only when necessary. One limitation of the RTS implementation for GPS and INS integration that should be noted is that when carrier phase ambiguity states are modelled, the INS will not have the advantage of using fixed ambiguity GPS measurements that may have been resolved if a full reverse filter implementation were to have been implemented. Therefore, to obtain the full advantage of post processing INS and carrier phase measurements, a conventional full forward and backward solution should be carried out. IMPLEMENTATION ISSUES The algorithms described so far in this paper have been implemented in the University of Nottinghams KinPosi GPS and INS integration software (for more information on KinPosi see, for example, Hide and Moore (2003) and Hide and Moore (2004a)). Our implementation of the RTS algorithm is slightly modified from the standard RTS algorithm implementation due to the feedback filter implementation described earlier in this paper. Using this approach it is the error in the INS that is modelled, so the predicted state at each epoch is always zero. Therefore, the above algorithm results in a series of smoothing corrections that must be applied to the navigation solution computed using the forward filter. Consequently, when the RTS algorithm is performed, the INS measurements are read in from the forward processing run, and corrected using the smoothed estimates. Furthermore, since the state transition matrix has to also be formed for the smoothing algorithm, variables such as the body frame IMU measurements also have to be stored (see, for example, Farrell and Barth (1999) for a description of the state transition matrix). Another factor that has to be considered is numerical stability as described in Shin and El-Sheimy (2002) who used Kalman filter smoothing for near real-time processing. This is largely because of the significant difference in the magnitude of the states that are estimated in the filter (for example position errors in metres are likely to be significantly larger than gyro bias error in radians per second). Therefore, problems can occur when computing the inverse of the predicted covariance in Equation 2. To reduce this problem, the method used by Shin and El-Sheimy is applied, where the units of the states are scaled before inversion so that they result in approximately equal magnitudes.

Another issue related to the particular implementation of the algorithm used in the KinPosi software is the use of multiple Kalman filter updates. In the following section GPS pseudorange and Doppler measurements are used. These measurements are considered to be uncorrelated and therefore they can be used to correct the INS using separate Kalman filter update steps. This approach is used to improve the efficiency of the filter, in particular by reducing the dimension of innovation covariance term that is inversed in the Kalman gain equation,
T T K k = Pk( ) H k ( H k Pk( ) H k + Rk ) 1

(4)

If pseudorange and Doppler measurements are simultaneously used to update the INS, this results in a matrix of dimension 2N (where N is the number of satellites) which has to be inverted whereas, when two separate Kalman filter updates are used, two separate inversions of dimension N are required. This results in a more efficient Kalman filter implementation. After each Kalman filter update the INS is corrected for the estimated error which also has the small further potential benefit of reducing linearization errors in the filter. This multiple update approach results in two separate state corrections which are applied to the INS. For the RTS algorithm, these state corrections are summed to form a single state correction, and the final covariance matrix after both updates is stored as the updated covariance. FIELD TRIAL To test the performance of tight integration combined with Kalman filter smoothing for GPS and low cost INS integration, a field trial was conducted in Nottingham, UK. A test route was devised to incorporate a range of conditions from relatively clear GPS conditions, to semiobstructed surroundings in suburban locations, to deep urban conditions where there is a significantly restricted sky view. The route is shown in Figure 3 with approximate areas shown for the different conditions encountered. The test route was driven twice and each loop took approximately 25 minutes. At the beginning of the trial, a short 5 minute alignment period was performed. Firstly, one minute of gyro measurements were averaged which were used as initial gyro bias estimates. This was then followed by a 5 minute alignment period where the vehicle was driven in figure-of-eight manoeuvres to resolve the roll, pitch and heading, and to estimate the initial gyro and accelerometer biases. The initial heading estimate was obtained from the velocity of the vehicle before it came to a halt before the gyro averaging calculation. Roll and pitch were initialized as zero.

reference points around the vehicle to compute the lever arm. This was applied in the KinPosi processing software. The measurements from the GPS and low cost INS were processed using the University of Nottinghams KinPosi GPS and INS integration software. A 15 state Kalman filter was used to model the INS errors with three states each for position error, velocity error, attitude error, gyro bias error and accelerometer bias error. The values used for the initial covariance and process noise in the Kalman filter were previously estimated from prior trials. Differential pseudorange and Doppler measurements were used to correct the INS using the Kalman filter algorithms described earlier in this paper. The University of Nottinghams Applanix POS-RS navigation grade INS and GPS integrated system was also installed in the vehicle to act as a reference for the low cost solution. The POS-RS uses a Novatel OEM4 GPS receiver integrated with a high accuracy Ring Laser Gyro Honeywell CIMU. The stated performance of the CIMU is shown in Table 1. The system is integrated using the loosely coupled algorithm using updates from GPS. The GPS updates were only used when the carrier phase ambiguities are fixed, with the high performance IMU used to bridge any long GPS outages without significantly reducing positioning accuracy. All data from the Applanix POS-RS is logged internally to a hard disk which is later used to post process the data. The Applanix POS-RS is used in the following analysis as a reference or truth for the low cost system. The Crossbow IMU was rigidly attached to the Honeywell CIMU, and the lever arm separation between the two IMUs was applied when outputting the post processed solution from the Applanix post processing software. Table 1 Crossbow AHRS400CA and Honeywell CIMU performance IMU Error source Accel Gyro (ms-2) (/s) Honeywell Bias < 1x10-7 3x10-6 CIMU Noise N/A < 1x10-7 Scale Factor 100ppm 10ppm Saturation 300 N/A Quantisation N/A N/A Crossbow Bias <0.3 <2 AHRS400 Noise 2.8 x10-2 9.5 x10-2 CA Scale Factor <1% <1% Saturation 20 100 Quantisation 2.5 x10-2 1.4 x10-2 Figure 5 shows the satellite availability during the trial. The figure shows that the number of satellites varies considerably from complete GPS outages to up to 9 satellites during clear conditions. The figure shows the number of satellites used by the Kalman filter. In the case

End Dense urban Start Semi urban Dense trees Clear sky

Figure 3 Vehicle trajectory The low cost INS data was collected using a Crossbow AHRS400CA IMU, the performance of which is described in Table 1. The AHRS sensor is approximately 5 years old and is considered to be of equivalent performance to IMUs that are currently available for less than $3000. The IMU was integrated with a Novatel OEM4 GPS receiver, although only L1 pseudorange and Doppler measurements were used which means that potentially, a much lower cost receiver could be used. The main disadvantage of low cost GPS receivers (of the order of a few hundred dollars) is the increase in measurement noise which should not significantly degrade the positioning performance of the integrated system. The Novatel GPS receiver was used differentially with a Leica GPS530 reference receiver located on the roof of the IESSG building. The maximum baseline during the trial was approximately 5km. The Crossbow IMU was located on the floor in the rear of the test vehicle. The lever arm separation between the IMU and GPS antenna was calculated by using a total station as shown in Figure 4. A prism was located on all of the

GPS IMUs

Figure 4 Total station survey for the estimation of lever arm separation between GPS receiver and IMUs

where only one satellite is available, the filter does not use the GPS observation since the filter uses a double differencing approach, hence any periods where only one satellite is available are shown as zero in the figure. Table 2 shows that during the trial there were less than 4 satellites available 43% of the time meaning that a three dimensional position solution is not available using GPS alone (this includes the initialization period that was conducted in a clear GPS environment). This highlights the problem of high accuracy positioning in urban environments where GPS measurements are infrequently available. Furthermore, during many periods where there are 4 or more satellites available, the dilution of precision value was large due to poor satellite geometry. The table also shows that using the double differencing approach, there are a further 31.9% of epochs where 2 or 3 satellites are available that can be used to aid the positioning performance of the INS.

occurred approximately 1000s (17 minutes) into the trial. These very large errors occur during an extended period of limited GPS measurements, and also because the filter has not had enough measurements to estimate the inertial sensor errors sufficiently. These statistics demonstrate that the loosely coupled GPS and low cost INS system is simply not suitable for use in dense urban conditions where there is a significantly reduced satellite availability. The loosely coupled integration filter has, however, been shown to provide good performance in, for example, marine conditions where there is constant GPS availability (see Hide and Moore 2004b). Table 3 also shows the navigation performance of the tightly coupled filter estimate for the same dataset. Here it is shown that there is a significant improvement in the estimation of the INS navigation states. The standard deviation of the horizontal position is reduced to less than 2m which is similar to the expected performance of a differential pseudorange GPS receiver operating in mixed conditions. Furthermore, the maximum error that occurs is 10.1m and 15.7m in the north and east axes respectively. However, the table also shows that the vertical position error is significantly degraded with a standard deviation of nearly 14m and a mean error of 2m. The vertical channel for low cost INS has generally been shown to provide good short term performance during clear GPS conditions, however, it is shown that vertical positioning performance is significantly degraded during operation in a dense urban environment. Table 3 Mean and standard deviation errors of GPS and low cost INS system for loose and tight coupled filters Loose coupled Tight coupled Mean
Position (m)

Figure 5 Satellite availability during Nottingham trial Table 2 Satellite availability during Nottingham trial # Sats 1 2 3 4 5 6 7+ % 11.3 9.7 22.2 21.8 12.0 7.4 15.6 TIGHT INTEGRATION RESULTS Table 3 shows a comparison between the navigation performance of the tightly coupled Kalman filter solution and a conventional loosely coupled Kalman filter. The errors were obtained by comparing the low cost solution to the reference trajectory provided by the post processed Applanix POS-RS data. Both loose and tightly coupled solutions were obtained using the KinPosi software, hence a direct comparison between the two algorithms can be made. The loosely coupled filter uses position and velocity updates from a separate GPS-only Kalman filter, hence it is not possible to update the INS when there are four or less available satellites. The table clearly shows that the performance of the loosely coupled filter is very poor. For example, the standard deviation of the north and east position errors is approximately 20m and 50m respectively. In fact, the maximum position error experienced during the trial was 147m and 598m which

s.d. 19.93 50.22 22.82 0.576 1.333 0.892 0.172 0.199 1.357

Mean 0.24 -0.02 -2.03 -0.008 0.004 -0.046 -0.376 1.005 0.348

s.d. 1.85 1.75 13.92 0.147 0.130 0.801 0.080 0.074 0.889

North East Down North East Down Roll Pitch Yaw

1.20 -5.30 -2.47 -0.009 -0.168 -0.076 -0.307 1.011 0.383

The table also shows that the velocity and attitude states result in improved estimation using the tightly coupled filter. The north and east velocity errors are significantly reduced by approximately 4 and 10 times respectively, however, again the vertical error is shown to provide relatively poor performance. The attitude errors are also reduced significantly using the tight coupled filter with an

Attitude ()

Velocity (m/s)

improvement demonstrated in all three angles. The roll and pitch errors are well estimated due to their effect on the gravity vector. However, it is shown that yaw error is relatively large with a standard deviation of approximately 0.9 for the tightly coupled filter. This is a common problem in low cost INS and is caused by the dynamics dependence of this state. For low cost INS, the inertial sensor errors need to be constantly updated using measurements in the Kalman filter. Therefore, when no measurements are available, the Kalman filter is unable to improve the error estimation. Furthermore, the ability to resolve some of the weakly observable states in the Kalman filter is also dependent on vehicle dynamics. In particular, as already mentioned, the heading error and z-axis gyro bias error is dependent on the presence of horizontal acceleration in order for the error to become observable through position and velocity updates. Figure 6 shows the z-axis gyro bias estimate from the loose coupled and tightly coupled filters. At the beginning of the dataset where there is good satellite availability, it is shown that the loose and tight integration filters result in similar error estimates. However, after approximately 400 seconds, the filter estimates begin to diverge.

continue to estimate the inertial sensor errors during limited GPS availability. This also increases the likelihood that the Kalman filter will have measurements available when the vehicle undergoes horizontal dynamic manoeuvres. KALMAN FILTER SMOOTHING RESULTS The previous analysis has identified that it is essential to use the tightly coupled Kalman filter for low cost INS integration in the urban environment. However, it has also been identified that if a real-time solution is not required, there is a significant potential benefit provided by applying a Kalman filter smoothing algorithm. Figure 7 shows the horizontal positioning error of the GPS and low cost INS system compared to the Applanix POS-RS solution for the tightly coupled forward and smoothed filters. The figure shows that the solution from the forward filter suffers from relatively large positioning errors of up to 16m. Again, the largest errors occur towards the beginning of the dataset during the first significant GPS outage periods which is thought to be as a result of the poorer estimation of the inertial sensor bias errors. Figure 7 also shows the positioning error of the Kalman filter smoothing algorithm. Here, it is clear that there is a significant improvement in the estimation of the vehicle position. The maximum error from the smoothed solution is 3.02m, which occurs during the longest period of reduced satellite availability. In fact, the accuracy is high enough that the error in the reference solution from the POS-RS (which will also be, to a lesser extent, degraded during this time) could affect this error estimate. The figure demonstrates that the combination of tight coupling and Kalman filter smoothing results in high performance positioning in all conditions.

Figure 6 Comparison of z-axis gyro bias estimation using loose and tight coupled Kalman filters In the loosely coupled filter estimate, it is shown that the error remains constant at various times throughout the processing run. These points occur during periods of obstructed GPS where the Kalman filter does not have any measurements to update the INS. However, during these same periods, it is shown that the tight coupled filter is able to improve the error estimation of the gyro bias even during periods where there are only 2 or 3 satellites in view. It is thought that the error estimation is improved since the attitude errors shown in Table X show a significant improvement over the results obtained from the loosely coupled filter. Therefore, it is interesting to note that the tightly coupled filter appears to be able to

Figure 7 Horizontal position error of GPS and low cost INS system during Nottingham trial Figure 8 shows a typical example of one of the relatively long periods of reduced satellite availability that occurred during the trial. The environment in which is data was

collected is shown in Figure 9. Figure 9 shows that the surrounding buildings are generally only 3 to 4 storeys tall, but they are close to the edge of the road. Furthermore, the road runs in a north-west to south-east direction which is significant at higher latitudes (Nottingham is located at approximately 53 latitude) where the majority of satellites are located towards the south. Figure 8 shows the number of satellites, and the horizontal positioning error. The figure shows that the maximum number of satellites during the 300 second period is 5, with the number regularly falling below 4. It also shows that there are a significant number of epochs where 2 or 3 satellites are available that are used by the tight coupled filter.

smoothed solution, it is shown that the maximum position errors occur in the middle of the period of reduced satellite availability as expected. The smoothed position error makes use of measurements either side of these periods which reduces the amount of time since the last full position update, and also makes use of the better error estimation that occurs from using the full set of data. SUMMARY Table 4 shows the position, velocity and attitude errors of the forward and smoothed Kalman filters. The results show that the smoothed filter provides a north and east position standard deviation of 0.66m and 0.55m respectively. The vertical error is also significantly reduced to 2.44m, which is an improvement of nearly 6 times. Further improvements are shown in the velocity and attitude estimation. The attitude errors are all improved by nearly a factor of 2 by implementing the smoother. The yaw accuracy is still relatively poor at 0.48, therefore for applications that require high accuracy yaw estimation, further aiding of the INS is required. The mean error of the attitude solution is relatively consistent between processing runs and this is thought to be a result of the misalignment between the Honeywell and Crossbow IMUs caused through the installation in the vehicle. Table 4 Mean and standard deviation errors of GPS and low cost INS system for forward and smoothed filters Tight coupled Tight coupled forward only and smoothed Mean
Position (m)

Figure 8 Typical example of positioning error during restricted satellite availability In the middle of the data there is a period of 103 seconds where there are less than 4 satellites (furthermore it is also worth noting that the HDOP and VDOP values were 1.7 and 5.6 respectively, just before this period, and 1.9 and 1.9 respectively just after this period). The figure shows that when 4 or more satellites are available, the position error for the forward filter is kept at the 1m level or better. When three satellites are available, the position error begins to grow at a slow rate, particularly for the first 20 seconds where there is no appreciable deterioration in accuracy. However, when only 2 satellites are available, the error growth is generally more rapid. For the

s.d. 1.85 1.75 13.92 0.147 0.130 0.801 0.080 0.074 0.889

Mean 0.32 0.16 0.57 0.000 0.000 0.001 -0.388 1.040 0.310

s.d. 0.66 0.55 2.44 0.061 0.047 0.586 0.051 0.049 0.481

North East Down North East Down Roll Pitch Yaw

0.24 -0.02 -2.03 -0.008 0.004 -0.046 -0.376 1.005 0.348

CONCLUSIONS This paper has demonstrated the performance of a tightly integrated GPS and low cost INS system in the urban environment. This paper has also shown the performance of the tightly integrated system when the data is post processed using a Kalman filter smoothing algorithm. It has been shown that with the grade of IMU used, the loosely coupled filter is unsuitable for use in environments where the number of satellites in view is regularly less than 4. The tightly coupled filter has been

Figure 9 Typical example of GPS environment during the Nottingham trial

Attitude ()

Velocity (m/s)

shown to provide horizontal position errors of less than 2m standard deviation with a maximum error of 16m. For applications where the data can be post processed, the horizontal position standard deviation reduces to less than 1m with a maximum error of 3m. Further improvements in velocity and attitude estimation have also been demonstrated. It is clear from the results discussed in this paper that GPS and low cost INS integrated systems can meet the performance levels required for a number of applications. In real-time, the navigation solution is suitable for applications such as vehicle navigation. The algorithms demonstrate a generalised integration algorithm that can be applied to all applications. However, improvements to the data processing algorithm could further be made for the vehicle application by, for example, fixing the navigation states when the vehicle is static. The benefit of post-processing the data was shown to be substantial (for example, the maximum horizontal position error was reduced by over 5 times). Post processing GPS and low cost INS data is generally not considered, however, it has been demonstrated that such a system is suitable for applications such georeferencing datasets for pollution and environmental analysis, or for surveying applications such as inventory management. Furthermore, in applications such as marine or aerial survey, these algorithms could be used to reduce the requirement for continuous GPS availability. For example, for UAV applications, the banking angle of the aircraft would not be critical. Further work to be undertaken at the University of Nottingham will be the testing of the low cost INS with a low cost GPS receiver. Such a receiver will be of the order of a few hundred dollars using space based differential corrections from a system such as EGNOS. This would result in an integrated system cost of less than $2500 using off the self components and the cost could potentially reduce significantly given the ability to batch fabricate MEMS components. The additional heading measurement derived from GPS velocity could also be incorporated into the filter when only considering the vehicle situation although this would not be suitable for vehicle performance testing where slip angle is of interest. Furthermore, when the vehicle is static, zero velocity measurements could also be used to update the INS. This would result in a genuine low cost positioning sensor that can work reliably in all conditions and environments. ACKNOWLEDGMENTS The authors would like to Scott Wilson Pavement Engineering for the use of their vehicle used to collect the data presented in this paper. The authors would also like

to thank Sam Waugh and Jock Souter for their help with the vehicle installation. REFERENCES Farrell, J. A., Barth, M., 1999. The Global Positioning System and Inertial Navigation. McGraw-Hill. Gelb, A. (Ed.), 1974. Applied Optimal Estimation. Analytic Sciences Corporation. Hide, C., and Moore, T., 2004a. Low cost sensors high quality integration. Proceedings of NAV/AIS04, Location and Timing Applications., London. Hide, C., Moore, T., 2004b. Performance of GPS and low cost INS integration in marine surveying." International Hydrographic Review 5((2)): 64-76. Hide, C., Moore, T., 2003. "Adaptive Kalman filtering for low-cost INS/GPS." Journal of Navigation 56(1): 143152. Nassar, S., Schwarz, K.P., El-Sheimy, N. (2004). INS and INS/GPS Accuracy Improvement Using Autoregressive (AR) Modeling of INS Sensor Errors. Proceedings of the ION 2004 National Technical Meeting (NTM 2004), San Diego, California, USA, January 26-28, 2004. Noureldin, A., Sharaf, R., Osman, A., El-Sheimy,, N., 2004. INS/GPS Data Fusion Technique Utilizing Radial Basis Functions Neural Networks, In proceedings of IEEE Position, Location and Navigation Symposium, Monterey, CA, 2004. Shin, E-H. and El-Sheimy, N., 2002, Optimizing smoothing computation for near real-time GPS measurement gap filling in INS/GPS systems, In Proceedings of the Institute of Navigation GPS 2002. Shin, E. H. and El-Sheimy, N., 2005, In-Motion Alignment of Low-Cost IMUs, European Journal of Navigation, 3 (1), pp 40-50., February, 2005 Talaya, J., Alamus, R., Bosch, R., Serra, A., Kornus, W., Baron, A., 2004, Integration of a terrestrial laser scanner with GPS/IMU orientation sensors, Proceedings of the Congress of the International Society for Photogrammetry and Remote Sensing (ISPRS), July 2004. Van der Merwe, R. and Wan, E. A., 2004, Sigma-Point Kalman Filters for Nonlinear Estimation and SensorFusion: Applications to Integrated Navigation. In Proceedings of the AIAA Guidance, Navigation & Control Conference, Providence, RI, Aug 2004.