Why Better Mapping Technology Is Critical To Autonomous Vehicles

Autonomous cars find the way to their destination using a number of critical technologies, including some version of a global position system and a central brain to interpret that and other data. But many of those technologies are not reliable or accurate enough today, and may not be for years to come.

There are numerous reports of vehicles missing their stop, or trucks being guided into alleys that are too narrow or underpasses that are too low. The data used to make those decisions is sometimes incomplete or confusing, and coverage for communications technology is often spotty. But a vehicle may need to make sudden adjustments based on that data, and fully autonomous vehicles need to be able to react in predictable ways even when they don’t have a pre-programmed solution.

This helps explain why the rollout of fully autonomous vehicles is taking far longer than initial predictions. Unlike human drivers, autonomous vehicles (AVs) need to learn how to drive, read maps, and arrive at the destination using real-time data and through interaction with the surroundings. AVs must be aware of problems that would cause accidents or cause them to get lost, and they need to be able to take contingency actions rather than just shutting down.

This is difficult when everything is reduced to ones and zeros. Sophisticated AI algorithms process input from a digital map and then plot the appropriate paths for the AV to follow. While human drivers typically use a GPS system and something like a Google map to go from point A to point B, an AV first uses a high-level, global map to determine its direction. Once within proximity of the destination, “localization” processing takes over. At this point, the AV will need to “see” its environment. Google has painstakingly created 2D and 3D maps by collecting the data using various methods. These include having vehicles drive around various areas equipped with high-definition cameras to capture street views. But for autonomous driving, AV needs higher precision down to the two- to three-centimeter level.

While most people are familiar with global positioning systems (GPS), these are subsets of global navigation satellite systems (GNSS), which include many other mapping and surveying applications. Many OEMs develop or acquire mapping technologies to support advanced driving assistance systems (ADAS). Technologies used for localization in AV include HD mapping, 3-D mapping, multilayered digital maps, and real-time mapping solutions.

Yet, consumer maps such as those provided by Google are insufficient for autonomous driving. OEMs that are developing AVs need to develop their own mapping technologies, or work with partners to come up with digital maps, because the types of maps AVs need are more complicated and have multiple layers that autonomous driving systems can leverage.

AV-tailored maps are high-definition and include detailed information about localization data such as lanes, road signs, speed limits, and more. Equipped with this data, AVs do not need to read and interpret every road sign. Rather, they can do a quick comparison for verification or even make decisions when signs are blocked.

For example, HERE Technologies developed a map solution that includes HD lane, HD localization, and road models. By combining the semantic data from satellite, cellular, Wi-Fi, and sensor technologies, an AV employing the HERE technology can correctly interpret road signs and poles, similar to what a human driver would do.

Mercedes-Benz’s DRIVE PILOT, in the new SAE Level 3 S-class, integrates the HERE map, and is said to be capable of effectively “seeing around the corners.” Without driver intervention, the DRIVE PILOT can control the speed and distance between vehicles, driving safely within the lanes. Lexus reports its SUVs can perform similar functions using multiple controls, but they require human intervention.

Fig. 1: Layer maps are used to provide detailed location information for autonomous driving. Source: HERE Technologies

NVIDIA is also a player here with its DRIVE Map, a multilayered mapping system consisting of many localization road signs, including lane dividers, road markings, road boundaries, traffic lights, signs, and poles. “Highly accurate maps are critical for automated and autonomous driving,” said James Wu, vice president and general manager of NVIDIA. “For humans, maps that are accurate to within a few meters are sufficient for providing turn-by-turn directions. For highly automated and fully autonomous driving, however, much greater precision and accuracy is demanded. Maps must be able to operate with centimeter-level precision for accurate localization, the ability for the AV to locate itself precisely on the road.”

One of the reasons is that AI makes decisions based on algorithms, and does not have human common sense. Over the past six years, NVIDIA has been building multiple localization layers of data with cameras, radar, lidar, and GNSS. Building on this large database, the DRIVE Map also crowd-sources from millions of passenger vehicles, continuously updating and expanding the map.

Creating and maintaining high-precision digital maps for autonomous driving is more complex than most people realize. Having a universal digital twin potentially can accomplish that goal, but the cost of maintaining such a project can be exorbitant.

To this point, Swift Navigation is taking a different approach to achieve the high-precision goal of two to four centimeters. Combining the signals from the GNSS, including GPS and the ground base stations, its solutions use a mathematical model to correct the errors created from the satellites, resulting in a high-precision positioning of the vehicle in motion.

One of the problems that automotive GPS/ GNSS receivers encounter is the loss of satellite signals in situations such as city driving with multiple high-rise buildings. To remedy that, Swift configures its system with STMicroelectronics’ Teseo GNSS receivers and a six-axis inertial measurement unit. When vehicles go through a downtown with high-rise buildings, or canyons and tunnels, the IMU detects and translates acceleration changes in position relative to the last satellite-detected position to overcome the temporary loss of satellite signals.

Fig. 2: Mathematical model combines GPS and ground station signals to achieve 2 to 4 cm map precision. Source: Swift Navigation

“Normal GNSS provides an accuracy of a few meters,” said Joel Gibson, executive vice president and general manager of automotive at Swift Navigation. “A high-precision solution should yield an accuracy of 2 to 4 cm. Additionally, keeping integrity risk or error rate to the minimum is also very important. As a comparison, HD cameras used in automotive have a typical error rate of 10^-3 per hour while the skylark solutions can keep it to 10^-7 per hour.”

The ultimate technology integration
The road to fully autonomous driving will be a long journey, and it will undergo integration of many different types of technologies. But along the way, drivers will enjoy the benefits of SAE Level 3 and Level 4 driving.

To reach level 5 requires more advanced AI, complex algorithms, multilayer precision maps, wireless connectivity of LTE, 5G, V2X, and more. All of these will need to work faultlessly with electronic control units (ECUs), sensors (cameras, radar, lidar, IR, ultrasound, and more). Some of these technologies are still in their early stages. Adding over-the-air (OTA) updates to the mix makes it even more challenging. Ultimately, safety is at the top of the priority list.

In designing AVs, it is critical to ensure driving safety when one or more of the above technologies fail. For example, 5G can provide very fast connections with very low latency, but an autonomous vehicle cannot solely depend on the connection for information. It must be able to be fully functional using automotive edge computing even without the network connection. Additionally, failover mechanisms need to be in place. Whether it due to a loss of signals, ECU failure, malware attack, or other hardware or software failures, AVs need to have a “soft landing” to safely stop and park. This is easy to do in city driving at speeds of 30 mph or less. It is a very different story at highway speeds.

“To increase safety and efficiency, AVs rely on multiple inputs for information,” said David Fritz, vice president of hybrid and virtual systems at Siemens Digital Industries Software. “These include cloud-based data from V2X, 5G, and GNSS. Additionally, AVs generate local data from various sensors. As autonomous driving moves from Level 2/3 ADAS to fully autonomous, complexity increases. Real-world data collection can be correlated with virtual model-based scenarios using digital twins. Leveraging the data in this way makes detecting complex operational anomalies possible, and will lead to higher AV safety and reliability. Additionally, constant learning and adaptation in almost real-time keeps the digital twin models trustworthy. These models will become more robust as more data are collected.”

Other concerns
At the same time, all the technologies available today, such as GPS, 5G, V2X, have to work together cohesively to achieve autonomous driving accuracy.

“First, cloud-based layer mapping technology using cellular and satellite have to work hand-in-hand with the vision technology using sensors like lidar to ensure localization accuracy,” said Charles Dittmer, senior product manager for wireless IP at Synopsys. “These technologies will continue to evolve. By 2025, there will be 10 million cellular V2X stations working together with the 5G macrocell towers. V2X is important for the level II and level III autonomous driving.”

All of this technology also has to be secure, particularly if a car is dependent upon an external map. OEMs and their supply chains are keenly aware of the cybersecurity challenges ahead.

“The number of cyber-attacks is in the tens of thousands per day across the globe,” observed Chris Clark, automotive software and security solutions architect at Synopsys. “In a traffic jam situation, malware attacks may result in minimal damages because the vehicles are not in motion. But it is a totally different story if the attacks occur while the car is traveling at higher speeds. Fatality may be the consequence. The auto industry is well aware of the cyber risks and working to deal with the risks proactively. For example, the World Forum for the harmonization of vehicle regulations enforces requirements to increase technology safety and security. The insurance industry also is looking at this issue, which could lead to increased vehicle insurance premiums. Potential traffic accidents are no longer limited to drivers’ fault. Automotive systems like ADAS will require new and novel ways of developing and testing to ensure security and safety at any speed.”

In addition, these vehicles need good-enough connections to work properly, and much of that depends on 5G, which at the millimeter-wave frequencies is prone to disruption and rapid signal attenuation.

“5G is still increasing its implementations of new technology in 3 categories — IoT, mobile, and automotive,” said Ron DiGiuseppe, senior automotive IP segment manager at Synopsys. “There are many 3GPP Releases that dictate 5G and future 6G technologies within the three categories. The V2X pieces are in their infancy, even within 5G, so 5G will take some time to be implemented, and 6G will take even longer. Development and adoption are taking place at a more rapid pace each year, but it still takes time. What’s exciting is that the innovations are mind-blowing. They will improve our lives in ways we may not have imagined yet, but they will have to work within the confines of safety. The Euro NCAP safety five-star rating is sought after by every automaker, which will include V2X in the safety requirements by 2025.”

The good news is that 5G cellular networks have many safeguards built in that take into consideration modern cloud deployments, and many networks are moving toward zero-trust architectures.

“5G has many security enhancements specifically targeted towards the transportation sector, including the early application of V2X, early warnings for anti-collision and traffic alerts, threat vectors such as bogus messages, Sybil attack, DoS, eavesdropping, and impersonate attack,” said Viet Nguyen, director of public relations and technology at 5G America.

That addresses some of the concerns, but there are many more. “The requirements set by OEMs and infrastructure constraints trickle down into semiconductor aspects,” said Frank Schirrmeister, vice president, solutions and business development at Arteris IP. “Safety and security aspects are critical, and privacy concerns also will play an important role, given all the data exchange across various boundaries. Observations made in 2011 on the EN-V regarding the design challenges in a complex system like that are huge, and offer great potential for more and improved design tools is still very accurate a decade later. Just think about laying out the network within the device and all the cross-talk effects. The protocol and software effects for networking within the vehicle, as well as between vehicles, are a definite challenge. The coordination of all the information necessary for driving and presenting it using a human-machine interface is a very complex task in itself. And, of course, bringing together all the mechanical and electronic effects will require complex cross-domain simulation.”

Conclusion
Improving accuracy is essential for autonomous vehicles. Getting lost or reaching a wrong destination is not an option, and bad data can cause accidents.

Many startups are developing new or improved mapping technologies. Other companies, including Intel, Qualcomm, and NVIDIA, are developing platforms for future automotive architectures promoting their own SoCs. And as autonomous driving moves from Level 3 to 4 or higher, mapping technologies will continue to evolve and improve.

Today, most of the OEMs are investing in autonomous driving and associated mapping technologies. This will be essential in the longer term, but it also is a competitive selling point in the near term.

But perhaps the biggest challenge for OEMs is integrating all of these technologies, both initially and over the lifetime of a vehicle. Software-defined vehicles, centralization of ECUs, mastering wireless technologies, and working with various sensors, will remain enormously complex and challenging. And while all of these new features can improve safety and convenience, there is still a long way to go.