It’s Time for 3D Mapping–Aided GNSS PAUL D. GROVES UNIVERSITY COLLEGE LONDON (UCL)
eal-time position accuracy, achievable in dense urban areas using low-cost equipment, is currently limited to tens of meters. If this could be improved to five meters Ror better, a host of potential applications would benefit. These include situation awareness of emergency, security and military personnel and vehicles; emergency caller location, mobile map- ping, tracking vulnerable people and valuable assets, intelligent mobility, location-based services and charging, augmented reality; and enforcement of curfews, restraining orders and From their beginning, GNSS technologies were known other court orders. to have limitations in certain operating environments, A further accuracy improvement to around two meters including anywhere that the low-power satellite would enable navigation for the visually impaired, lane-level road positioning for intelligent transportation systems, aerial signals could be blocked or dispersed. This has given surveillance for law enforcement, emergency management, rise to a variety of integrated solutions designed building management, newsgathering, and advanced rail sig- to extend and improve the performance of GNSS naling. under adverse conditions. Among these solutions This article explores how 3D mapping can be used to in recent years is the use of three-dimensional achieve a “step change” in real-time GNSS performance in maps to support several new techniques, including dense population centers, including urban canyons. “shadow-matching,” which seek to turn the disruptive The Problem effects of buildings in urban areas into a method Buildings and other structures block many GNSS signals in for actually enhancing position determination. urban environments. When obstructions block satellite sig-
50 InsideGNSS SEPTEMBER/OCTOBER 2016 www.insidegnss.com nals that cross the street at right angles, allowing only along- hardware cost. However, this approach does not mitigate NLOS street signals to reach a receiver, signal geometry and position- reception errors at all. ing accuracy decline. In extreme cases, where the ratio of the Multi-constellation receivers have significantly improved building heights to the street width is very high, the number the availability of GNSS positioning in dense urban areas. of directly received signals may be insufficient to compute a However, accuracy remains a problem. In hostile signal prop- position solution. agation environments such as dense urban areas, the stand- Another well-known signal propagation effect occurs when alone, single-epoch positioning accuracy of GNSS is degraded buildings, walls, vehicles, and the ground reflect GNSS signals. to an average of about 25 meters. In the worst cases, position Metal, metallized glass, and wet surfaces are particularly strong errors can exceed 100 meters. reflectors. Where the direct line-of-sight (LOS) signal is blocked Filtered or carrier-smoothed positioning algorithms can and only a reflected, or non-line-of-sight (NLOS), signal is perform better, but these require an accurate position solution received, a ranging error occurs that is equal to the additional for initialization. They then need a sufficient number of good- path taken by the NLOS signal. quality GNSS signals to maintain solution accuracy and, hence, Strong NLOS signals can be difficult to distinguish from the ability to reject poor measurements using innovation filter- direct LOS signals, particularly using a smartphone, which ing. In practice, this can be difficult to achieve in challenging has a linearly polarized antenna. Reflected signals can interfere urban environments. with reception of the direct LOS signals, a phenomenon known Outdoor Wi-Fi positioning is accurate to around 25 meters, as multipath interference or more simply, multipath, because the while other radio positioning techniques either offer poorer signal is received via multiple paths. Figure 1 illustrates both accuracy or require expensive dedicated infrastructure. Dead- phenomena. reckoning techniques, including wheel-rotation sensing and GNSS user equipment can minimize ranging errors due to pedestrian dead-reckoning using step detection, require accu- multipath interference by exploiting receiver signal processing rate initialization, and their position accuracy then decreases technology at the expense of increased power consumption and as distance measurement errors accumulate. www.insidegnss.com SEPTEMBER/OCTOBER 2016 InsideGNSS 51 3-D MAPPING-AIDED GNSS
most needed. Conventional stand-alone GNSS positioning works well enough in low-density areas. The final ingredient is advanced positioning algorithms that take advan- tage of 3D mapping. Many algorithms have been developed. Some use pseudo- FIGURE 1 Multipath interference and NLOS reception ranges while others use SNR measure- ments with varying trade-offs between
If we are to achieve low-cost, real- (C/N0). Pseudorange rate (Doppler) and performance and processing load. Over time accuracy of five meters or better carrier phase can also be useful. Sur- the past five years, results have been in challenging urban environments, a vey receivers have always provided this published from more than 10 different whole new approach is needed. Current- information, but obtaining these data research groups spread across Europe, ly, 3D mapping–aided GNSS presents a from consumer receivers has histori- North America, Japan, and the Middle great opportunity. cally been problematic. Today, however, East. some low-cost GNSS receivers generate The Opportunity pseudorange and SNR measurements What Can Be Done? To implement 3D mapping–aided GNSS, from all GNSS constellations, providing 3D mapping can be used to provide we need three things: measurements, access to this data through the applica- three types of information for aiding mapping, and algorithms. In dense tion programming interface (API) on GNSS positioning. The first, and sim- urban areas, the more satellites that are smartphones and tablets running the plest, is the terrain height. Next are the available for use, the better. Therefore, Android Nougat operating system that predictions of which satellites will be the return of GLONASS to full opera- have a compatible GNSS chipset. directly visible at a given position and tional capability in 2011, followed by the The second ingredient, 3D mapping, time, i.e., which lines of sight are blocked wide availability of GLONASS-capable can prove expensive when it comes to by buildings and which are not. receivers presented a step-change in sig- highly detailed 3D city models. How- Finally, 3D mapping can be used to nal availability. In 2016, with the Galileo ever, simple block models, known as predict signal reflections, including the satellite deployment well under way and level-of-detail (LOD) 1, are sufficient for path delay of the reflected signal with the extension of BeiDou from a regional most 3D mapping–aided GNSS imple- respect to the direct signal and, poten- to a global system, we are in the midst of mentations. (Figure 2 shows an exam- tially, its relative amplitude. However, a second step-change. ple.). The free, editable OpenStreetMap amplitude is difficult to predict because To implement advanced positioning provides building maps for major cities 3D mapping data doesn’t include infor- algorithms, we need access to the “raw” worldwide, much of it three-dimen- mation on building reflectivity at GNSS measurement data, namely, the pseudo- sional (This can be viewed online using, frequencies, which can be quite different range and signal-to-noise ratio (SNR) for example, the F4 Map demo, noting from optical reflectivity. or carrier-power-to-noise-density ratio that a default height is assumed for 2D Terrain height-aiding is useful buildings.) because, by constraining the posi- Data is avail- tion solution to a surface, it effectively able from national removes a dimension from the posi- mapping agencies tioning problem. In open areas, terrain with various terms height-aiding only improves the vertical and conditions, position solution (as one might expect). and private com- However, in dense urban areas where the mercial companies signal geometry is poor, it can improve such as Google and the horizontal accuracy by almost a fac- Apple also hold tor of two. large amounts of We can use satellite visibility predic- 3D mapping data. tions to aid both conventional ranging- Although such digi- based GNSS positioning and a new class tal map coverage is of SNR-based positioning algorithms far from universal, known as shadow matching. Each GNSS FIGURE 2 LOD 1 3D model of City of London. (Data from Ordnance it tends to be avail- signal is predicted to be directly visible Survey Mastermap, Crown Copyright and database rights 2016. able in dense urban in some areas and blocked (shadowed) Ordnance Survey (Digimap Licence) areas where it is in other areas.
52 InsideGNSS SEPTEMBER/OCTOBER 2016 www.insidegnss.com between the measured and predicted pseudoranges (assum- ing LOS propagation). Different assumptions about the error distribution can then be made at various candidate positions — according to which signals are predicted to be LOS or NLOS at each position. Thus a symmetric error distribution can be assumed for LOS signals and an asymmetric distribution for NLOS signals, with the scoring adjusted accordingly. The candidate positions may be distributed in a regular grid or semi-randomly (as in a particle filter). Terrain height-aiding is used to associate a height with each horizontal coordinate, enabling the search space to be limited to two dimensions. FIGURE 3 Principle of shadow matching University College London (UCL) has combined this hypoth- esis-scoring ranging algorithm with shadow matching, which Shadow matching assumes that the user is in one of the has reduced the root mean square (RMS) horizontal position- directly visible areas if the received SNR is high and in one ing error in dense urban areas from 26 meters to 3.9 meters of the shadowed areas if the SNR is low or the signal is not using data from a consumer-grade GNSS receiver — a sixfold received at all. Figure 3 illustrates the general principle. Repeat- improvement. Figure 4 shows the across-street and along- ing this calculation for each GNSS signal enables one to reduce street positioning errors obtained at 18 sites in central London, the area within which the user may be found. depicted in Figure 5. In practice, the SNR distributions of direct LOS and 3D mapping can also be used to predict reflected signals. NLOS signals can overlap, particularly when using a smart- For shadow matching, this potentially provides more informa- phone antenna. Furthermore, real urban environments and tion from which to predict the SNR measurements. For rang- signal propagation behavior are more complex than can be ing, this can be used to predict which direct LOS signals are represented using 3D mapping. Therefore, a practical shadow- subject to multipath interference, enabling us to adjust their matching algorithm works by scoring a grid of candidate posi- weighting or assumed error distribution within the position- tions according to the degree of correspondence between the ing algorithm. satellite visibility predictions and the SNR measurements. This Where the path delay of the reflected signals is predicted, enables inaccuracies in the process to be treated as noise so that NLOS reception errors may be corrected, enabling NLOS sig- a correct position is still obtained provided there is sufficient nals to contribute to an accurate position solution. However, “signal.” Due to the building geometry, shadow matching is normally more accurate in the cross-street direction than the along-street direction, which complements ranging-based GNSS position- ing in dense urban areas. Shadow matching enables users to determine which side of the street they are on in environments where other positioning technologies do not. Satellite visibility predictions can be used to aid ranging- based positioning in a number of different ways. Where the position is already known to within a few meters, we can pre- dict which signals are NLOS with reasonable accuracy and sim- ply exclude them from the position solution (assuming there are sufficient direct LOS signals). Otherwise, which signals are directly visible depends on the actual position of the user. A simple approach is to determine the proportion of candidate positions at which each signal is predicted to be directly vis- ible. This information is used to weight each measurement within the position solution and to aid consistency checking. This approach typically improves the positioning accuracy by 20–25 percent and can handle initialization errors of about 100 meters. To make the best use of satellite visibility prediction, a con- ventional least-squares (or extended Kalman filter) position- FIGURE 4 Across-street and along-street positioning errors ing algorithm should be replaced by an algorithm that scores across 18 sites in central London using conventional and 3D mapping–aided GNSS candidate position hypotheses, according to the difference www.insidegnss.com SEPTEMBER/OCTOBER 2016 InsideGNSS 53 3-D MAPPING-AIDED GNSS
rithms for predicting GNSS signal propagation using 3D map- ping are much more computationally intensive. Performing ray-tracing of the of a signal’s path for 20 satellite positions and 1,000 user positions can take several minutes of central processing unit (CPU) time. This problem has two solutions. The first is pre-computa- tion. For predicting satellite visibility at street level, 3D building boundaries — representing the surfaces and edges of a struc- ture — can be computed for a one-meter grid of candidate posi- tions. Each building boundary comprises the elevation thresh- old below which satellite signals are blocked for each azimuth. Satellite visibility can therefore be predicted very quickly simply by comparing the satellite elevation with the building bound- ary elevation at the appropriate azimuth. The main drawback to this approach is that the building boundary data can take up more space than the original 3D mapping. The building boundary approach can also potentially be extended to predicting whether or not a reflection is received. The path delays of reflected signals could be pre-computed, but this would create a very large amount of data; so, it may not be practical. FIGURE 5 Locations of test sites in central London (background image from Google Earth) The second approach is to use a graphics processing unit (GPU). A GPU is designed for parallel processing, which accurate correction of NLOS errors requires an accurate posi- enables us to consider multiple candidate user positions simul- tion solution, a “chicken and egg” problem. If we already know taneously. Using projection, the visibility of 20 satellites over the position to within a few meters, alternate computation of a 100 × 100-meter position grid may be computed in about a the position solution and NLOS corrections may be iterated second. until they converge. For larger uncertainties, we will need To compute the path delays of reflected signals, ray tracing multiple starting positions to ensure convergence. Another can be run on a GPU, enabling computation of paths from 20 approach, known as the “urban trench” method, incorporates satellites to 100 candidate positions within a second. Operat- the reflecting surfaces within the positioning algorithm. How- ing the GPU on a mobile device increases power consumption ever, this only works where we can determine which surface because GPUs typically consume about twice the power of reflects which signals. CPUs. However, the power consumption of the GNSS receiver A more powerful approach adds NLOS error prediction to chip is also significant. Because of their operating design, GPUs positioning by scoring candidate position hypotheses. NLOS may also speed up the generation of building boundary data. corrections are computed for each candidate position. Both Distribution of 3D mapping data is another important fac- grid-based and particle-based methods have demonstrated tor to consider. Three options for distributing these data include positioning accuracies within two meters. However, they are pre-loading, streaming, and server-based positioning, as shown computationally intensive. in Figure 6. In all cases, a binary data format should be used to These different approaches should not be treated as compet- minimize the capacity required. itors. SNR-based and pseudorange-based algorithms are clearly complementary. Better performance can be achieved by using both and combining the results. Similarly, a computationally efficient algorithm, oper- ating over a wide search area, can be used to initialize a higher-resolution, computationally intensive algo- rithm operating over a smaller area.
How Practical Is It? For 3D mapping–aided GNSS to be practical, the pro- cessing load and data storage or transmission load must match the intended applications. The algorithms for determining position will run relatively efficiently FIGURE 6 Pre-loading, streaming and server-based system architectures on a PC, tablet, or smartphone. However, the algo-
54 InsideGNSS SEPTEMBER/OCTOBER 2016 www.insidegnss.com The terrain height data are easiest to which should be no larger than 100 ly, it can operate in real-time with pre- handle. A five-meter grid spacing is suf- meters by 100 meters, considering only computed building boundary data, or ficient, corresponding to 40,000 points- outdoor locations. potentially GPU-based projection from a per- square kilometer. Twelve bits is Only azimuths corresponding to the 3D building model. Thus, 3D-mapping- sufficient to describe the relative height current set of GNSS satellites are need- aided GNSS positioning is a practical of a point within a tile. Four bytes are ed, which reduces the amount of data proposition. needed for the height of each tile’s origin required to 90 kilobytes without com- Other research groups have dem- with respect to the datum. Thus, about pression. The pre-computed path delays onstrated higher accuracy (around two 60 kilobytes are needed to represent a would comprise one value per satellite, meters) using methods that predict square kilometer of mapped urban ter- per candidate location, requiring about NLOS path delay using 3D mapping. rain; so, one gigabyte of storage could 125 kilobytes for 10-bit values. Third- These are too computationally inten- accommodate map data representing generation (3G) mobile download speeds sive to operate in real-time over a large about 17,000 square kilometers, Even are more than 500 kB/s (4 Mbit/s). search area on a mobile device. However, more can be stored using data compres- Therefore, streaming of mapping data is they could be used as the final step in sion. practical, and substantial data buffering a multi-stage positioning process, fol- For 3D modeling of a city’s build- could be accommodated to bridge any lowing on from the UCL algorithms. ings, about 500 bytes might be used gaps in communications coverage. With further research, further viable to describe a building to LOD 1, and a The final option is to calculate the approaches are likely to emerge. square kilometer in a dense urban area position solution on a remote server. might typically contain about 1,000 This requires uploading of the GNSS The Way Forward buildings. Thus, about 100 kilobytes pseudorange and SNR measurements The GNSS research community has per square kilometer is needed; so, one from the mobile device to the server, demonstrated that 3D mapping–aided gigabyte of storage could accom- GNSS can vastly improve posi- modate about 2,000 square kilo- tioning accuracy in dense urban meters of data, or enough to cre- areas and is practical to imple- ate 3D maps for about two cities The GNSS research community has ment. Current algorithms could — again, more with compression. be deployed in consumer loca- Building boundaries require demonstrated that 3D mapping–aided tion-based services right now. a lot more data. To achieve a one This could operate as an enhance- degree precision, requires about GNSS can vastly improve positioning ment to server-side AGNSS using 300 bytes per building bound- current interfaces and protocols. ary. Assuming that about half the accuracy in dense urban areas and is A service provider would have to space in a city is outdoors (build- invest in additional hardware to ing boundaries are not required practical to implement. run the positioning algorithms for indoor locations), a 100 × for multiple users simultaneously 100–meter tile would require and store the 3D mapping. 1.5 megabytes of data without 3D mapping–aided GNSS compression. So, one gigabyte of stor- and then downloading the position algorithms could also run on a mobile age would only accommodate about 7 solution. This imposes a minimal com- device, potentially within an app, pro- square kilometers of data, perhaps 70 munications load and could employ cur- vided the device has a compatible API square kilometers with compression. rent assisted-GNSS (AGNSS) protocols; and GNSS receiver chip. In this case, a Thus, pre-loading may be practical if so, it would be compatible with all cur- service provider would be required to the 3D mapping is used directly, but is rent mobile devices. However, it needs store the 3D mapping data and down- unlikely to be if building boundaries are continuous communications coverage. load it to users on demand. Clearly, used. Pre-loading of pre-computed path A server would almost certainly use pre- either architecture needs a viable busi- delays would not be practical. computed building boundaries and pos- ness model to fund it. One possible Moving on to streaming, if 3D map- sibly pre-computed path delays as well, driver is enhancing a mobile navigation ping is used directly, buildings within a depending on the size of the user base. app’s user experience to attract more ~300-meter radius of the predicted user UCL’s combined shadow-matching customers. position should be downloaded as they and GNSS-ranging system, which is Many specialist applications — such could potentially affect signal reception. accurate to about four meters, uses pre- as navigation for the visually impaired, This download would require about 150 dictions of terrain height and satellite situation awareness of emergency, secu- kilobytes of data. For building bound- visibility based on the 3D mapping, but rity and military personnel and vehicles; ary data, only the search area is needed, not path delay predictions. Consequent- emergency caller location; and tracking www.insidegnss.com SEPTEMBER/OCTOBER 2016 InsideGNSS 55 3-D MAPPING-AIDED GNSS
vulnerable people and valuable assets — project EP/L018446/1, “Intelligent [13] Suzuki, T., and N. Kubo, “Correcting GNSS Mul- tipath Errors Using a 3D Surface Model and Particle could also make use of current 3D map- Positioning in Cities using GNSS and Filter,” ION GNSS+ 2013, Nashville, Tennessee USA, ping–aided GNSS algorithms. With a Enhanced 3D Mapping.” September 2013 much smaller user base than consumer The photos on the opening pages of [14] Suzuki, T., and N. Kubo, “GNSS Positioning with applications, much less server infra- this article were taken by Dr. Adjrad and Multipath Simulation using 3D Surface Model in structure would be required to get them former UCL students Nachuan Li and Urban Canyon,” ION GNSS 2012, Nashville, Tennes- operational. Christelle Mekemlong Lando. see USA, September 2012 The scientific community is another [15] Wang, L., “Investigation of Shadow Matching Additional Resources for GNSS Positioning in Urban Canyons,” Ph.D. The- potential early adopter of 3D mapping– sis, University College London, 2015. Available from aided GNSS. Many research groups use [1] Adjrad, M., and P. D. Groves, “Intelligent Urban
56 InsideGNSS SEPTEMBER/OCTOBER 2016 www.insidegnss.com