Delivering continuous lane-accurate positioning for autonomous driving
01 October 2019
Today’s state-of-the-art, single-band GNSS receivers fulfil accuracy requirements for V2X, ADAS & autonomous driving in ideal open-sky conditions. But to reliably serve these use cases everywhere, all the time, they need to overcome their limitations in urban & other challenging environments.
This article was originally featured as the cover story in the October 2019 issue of EPDT magazine [read the digital issue]. Sign up to receive your own copy each month.
Here, Alex Ngi, Product Manager, Product Strategy for Dead Reckoning at positioning & wireless technology specialists, u-blox demonstrates how this can be achieved with a multi-band, RTK, dead reckoning system using GNSS correction services and a dynamic vehicle model.
Satellite-based positioning plays a unique enabling role for V2X applications and advanced driver-assistance systems (ADAS), including autonomous vehicles. It is the only technology capable of determining a car’s absolute position in real time. It works independently of maps, cameras and landmarks. And because its fundamental operating principle is completely unrelated to other sensing technologies (such as LIDAR, cameras or ultrasound) used in self-driving vehicles, satellite-based positioning offers a core foundation and backbone for the multi-sensor network that no other technology can provide.
Today, global navigation satellite system (GNSS) receiver technology is overcoming its own inherent limitations, one after another. Accuracies are coming down to just a few tens of centimetres, with convergence times (the time it takes a receiver to reach a predetermined accuracy level after signal interruption and subsequent reacquisition) in seconds. Latencies (the time between the position measurement and when the device reports this position to the network) are approximately 10 milliseconds. Position updates can be delivered at well above 10 Hz. In addition, with further technological enhancements, positioning can extend deep into urban canyons, under multi-level road structures, and other challenging scenarios.
In short, GNSS is finally technologically ripe for the era of V2X and ADAS applications.
Not all the evolution took place in the GNSS receiver, though. Moore’s Law shrunk the hardware required to the size of a miniature chip suitable for mass market, portable, low-power devices. Ubiquitous wireless internet connectivity enabled GNSS correction services that minimise the ionospheric effect on GNSS accuracy, the dominant source of GNSS errors. Finally, national and international investments in the space sector gave us new constellations of satellite systems designed for innovative applications. These provide the key advantage of additional satellites available (visible) to the receiver.
These advances will put vehicles on our roads equipped with the latest generation of multi-band, multi-constellation GNSS receivers capable of delivering sub-metre accuracies, even down to a few tens of centimetres, depending on the requirements of the application.
But it isn’t just about improved positioning accuracies. Low latency is another critical requirement in emerging applications, such as vehicle-to-everything (V2X) communication. In V2X, vehicles ‘talk’ to each other and with roadside infrastructure using wireless messages, passing on warnings and information about manoeuvers and negotiating priority at intersections, when merging and overtaking.
At best, long latencies might be a nuisance, leading to unnecessary braking and acceleration, reduced efficiency of truck platoons, and a drop in passenger comfort. But at worst, they can be deadly, particularly on motorways, where cars travel the length of a car in just a tenth of a second. The ETSI (European Telecommunications Standards Institute) standard on V2X communication requires latencies below 100 milliseconds at the system level for most use cases.
Advanced sensor fusion filter with low convergence time
For ADAS, V2X and ultimately for autonomous driving to be viable, GNSS receivers must robustly deliver lane-accurate positioning, even in challenging environments. When satellite signals are temporarily obstructed, they need to recover the high precision position solution in seconds, which can be achieved by combining a number of complementary elements that are implemented in a single sensor fusion filter presented in Figure. 2.
Multi-constellation, multi-band GNSS receivers: The increase in the number of global GNSS constellations from one (GPS) to four (GPS, GLONASS, BeiDou, Galileo) means that receivers ‘see’ more satellites from any given location. This more than makes up for the increase in number of satellites receivers need to unambiguously determine their position: four for a single constellation, roughly seven when three constellations are used (in order to compute the time differences between the constellations, which have inherently different time references from each other).
In addition to more satellites, multi-band GNSS receivers can combine signals at different frequencies, each with application-specific benefits. For example, simultaneously processing two signals from different frequencies effectively removes up to 99.9% of the ionospheric error. Another technique, called geometry-free combination, helps in detecting cycle slips in the carrier phase. All these techniques are only possible with multi-band receivers.
Integrated real-time kinematic (RTK) algorithms: While standard-precision GNSS receivers track the code phase of GNSS signals from at least four GNSS satellites to trilaterate their position, high-precision GNSS receivers track the high-frequency carrier phase. In order to resolve carrier-phase ambiguities, high-precision GNSS receivers make use of real-time kinematic (RTK) algorithms, which in some cases are integrated into the GNSS receiver module. The RTK algorithms make extensive use of correction data delivered over a wireless connection. For the automotive market, cellular and satellite L-band based communications are ideally suited. In addition to saving data transmission costs, L-band receivers can receive RTK correction data via satellite, even in rural areas, where cellular connections are poor or not available at all.
Broadcast GNSS correction services: GNSS correction service providers constantly estimate GNSS signal errors by monitoring them from a network of base stations. Precise point positioning (PPP)-RTK services, for example, compensate for satellite clock, orbit, signal bias, global ionosphere, and regional ionosphere and tropospheric effects. Ideally, corrections should be valid over large regions, such as the continental United States, and have minimal bandwidth requirements. While legacy services send out a tailored correction stream to individual users based on a rough position estimate, modern service providers adopt a more scalable approach, broadcasting the same dynamic GNSS error model to all their users.
In addition to increasing GNSS receiver accuracy, high quality correction data shortens the time it takes for the receiver to converge to a precise position estimate. This is critical in normal driving environments that include overhead obstructions such as overpasses, highway signage, trees and bridges, which can momentarily interrupt GNSS signals.
Inertial sensors and sensor fusion: Inertial sensors have been used to augment GNSS receivers for several years now. By enabling dead reckoning (DR), they allow vehicle positioning systems to bridge the GNSS gaps that receivers experience in tunnels, parking garages and other challenging, yet common environments. By fusing data gathered by the individual components of the inertial measurement unit (IMU), the positioning module can continue to deliver a position estimate in obstructed environments, where the GNSS signal cannot penetrate.
Inertial sensors and sensor fusion help the positioning solution retain information about position and velocity when GNSS signal reception is shortly interrupted. This shortens the reconvergence time – in other words, the time it takes to solve carrier phase ambiguities when satellite signals become available again, compared to GNSS-only solutions.
In-vehicle sensors: Incorporating data from in-vehicle sensors, such as the wheel-tick sensor, further enhances the performance of the dead-reckoning solution. The algorithms can reject position changes reported by GNSS inaccuracies (induced by signal obstructions), simply by knowing that the wheel did not move. The velocity reading, weighted with wheel-tick sensor data, is more accurate than a system relying on integrating the noisy accelerometer measurements. Furthermore, sensor calibration to determine distance travelled per revolution is performed continuously, and forms a virtuous cycle able to account for winter and summer minute tire changes.
Dynamic model: A dynamic vehicle model of the vehicle limits the effect of measurement errors on the reported position. The model assumes that the vehicle does not slide laterally, jump vertically or accelerate in any unreasonable way. All GNSS measurements are checked for plausibility before being used in the navigation filter.
Quantifying performance in tunnels
Quantifying the performance of the aforementioned solution in tunnels remains challenging. For one, the primary error source is sensor biases, and this tends to accumulate as they are integrated to derive the velocity (accelerometers) and the attitude (gyros) of the vehicle, due to predominantly stochastic, rather than systematic phenomena. To properly characterise their effect, a statistically relevant number of tunnels need to be driven and data collected. Second, there is no obvious ‘true’ position with which to compare measurements. Ideally, an alternate positioning technology, based on a completely different technology, is used inside these tunnels as a reference that is unaffected by obstruction of the sky. Finally, even expensive inertial sensor-based reference systems drift to some extent.
Rather than testing the setup against a truth system in actual tunnels, we first created virtual tunnels using data collected under open sky conditions. To do this, we ‘unplugged’ the GNSS receiver to simulate a GNSS outage, forcing the system to navigate in dead-reckoning mode. This let us compare the performance of the inertial measurement unit (IMU) against a high-end truth setup. Logging the readings from the dead-reckoning solution and the high-end reference GNSS receiver provided us the necessary data to emulate tunnels of different lengths on sampled portions of a given dataset. It was a simple trick that delivered a set of test runs large enough to qualify performance in a way that is statistically significant.
In the graph shown in Figure 3, analysing the data from 1758 outages generated from 31 test runs, we determined that our positioning error over distance travelled was approximately 2% in dead-reckoning mode. In other words, for each kilometre travelled, the horizontal positioning error grew by 20 meters on average. It’s worth noting that that the performance of the IMU strongly affects the tunnel test results. In our setup, we used a standard IMU with average performance, rather than a high-end one.
Tried and tested on asphalt
The tunnel simulations were only part of a broader series of device tests. To validate that the combination of technologies outlined – in other words, combining a multi-band, multi-constellation GNSS receiver with built in RTK algorithms, broadcast GNSS correction data, an IMU for dead reckoning, an external wheel-tick sensor and a dynamic vehicle model – reliably delivers lane-accurate positioning, we also tested it in a number scenarios of varying complexity. Because of the stochastic nature of GNSS and IMU errors, individual test runs can either over- or underperform compared to the results presented in Table 2.
During a recent drive on a highway, largely in open sky conditions – our least challenging test case – we confirmed that our solution delivers 100% availability and is accurate to within 5.8 centimetres 50% of the time. The horizontal velocity component was accurate to 0.02 km/h 68% of the time.
During our test, we registered a split between an RTK fix (integer ambiguities of the carrier phase resolved), RTK float (integer ambiguities of the carrier phase unresolved) and dead reckoning of 82% to 14.8% to 3.1%. Overall, the accuracy of the solution was improved tenfold over existing single-band receiver technology. It must, however, be cautioned that absolute comparisons of RTK fix and float may be misleading. For a given receiver, it is a good indication of relative level of difficulty between different test tracks while evaluating the accuracy achieved. It is a less useful figure of merit comparing two receivers.
Test results for open sky conditions on a highway and typical urban areas in Paris showed outstanding performance improvements compared to single-band, non-RTK setups. In a worst case urban canyon scenario, carried out in the La Défense district in Paris, the performance continued to exceed the requirements for V2X applications. Even though the GNSS receiver was unable to fully resolve the integer ambiguities of the carrier phase, CEP68 was about 1.1 metres (in this context, a CEP68 – short for circular error probability of 68% – of 1.1 metres means that 68% of readings are within 1.1 metres of the GNSS receiver’s true location on a two dimensional surface), with the solution accurate to at least 1.7 metres 95% of the time. This scenario shows clearly how the technology used can boost positioning performance in most challenging urban environments.
Finally, we tested the performance in a two-kilometre tunnel in Gothenburg, Sweden, where our solution performed better than in our extensive simulations. Drift was 50% lower than expected, at 1% of distance travelled. Furthermore, convergence back to lane-level accuracy took only two seconds. This rapid convergence was due to a combination of factors, including multi-frequency GNSS receivers, GNSS correction services and a relatively accurate estimate of the position inferred from dead reckoning. Clearly, lane-accurate positioning is not maintained in long tunnels. In such scenarios, however, highly automated and driverless vehicles can make up for the loss of accuracy using complementary positioning technologies.
An obvious value-add for automotive GNSS
In conclusion, positioning solutions combining multi-band, multi-constellation GNSS receivers with built-in RTK algorithms, broadcast GNSS correction data, an IMU for dead reckoning, an external wheel-tick sensor and a dynamic model are able to provide continuous lane-accurate positioning, even in the most challenging environments. This can be further enhanced by fusing the information from other vehicle sensors, such as cameras and radars. This will contribute in making our transport systems safer, more comfortable and more efficient. With this combined approach, GNSS technology augmented with dead reckoning is now ready for advanced automotive applications.
We found the solutions outperform existing technology tenfold in terms of accuracy. Continuous service in urban environments is achieved by a powerful combination of multi-band, multi-constellation GNSS receivers capable of maximising satellite visibility in partly obstructed scenarios, dead reckoning to bridge short gaps in GNSS reception, and GNSS correction services for fast reconvergence from short GNSS interruptions. Given their accuracy and global coverage, and the fact that GNSS provides the only absolute truth for position and time, advanced automotive applications are bound to benefit from integrating these technologies.
Contact Details and Archive...