GET THE PDF VERSION
Tired of scrolling? Download a PDF version for easier offline reading and sharing with coworkers.
I. The case for a smart AI-powered safety network
An innovative approach to enabling safer roads in New Zealand through Emerging Tech.
The potential of maturing Artificial Intelligence (AI) and Internet-of-Things (IoT) technologies to enable rapid improvements in road safety has been largely overlooked. On one hand, roading spend remains focused on traditional physical infrastructure, while on the other, technologists are distracted by a space race to build autonomous vehicles. Autonomous technology is difficult and it will likely take 20-30 years to mature and substantially replace existing fleets. Unaddressed, this gap will result in a lot of unnecessary social harm on our roads.
What can we do in the meantime?
Fortunately, there is an opportunity to utilise existing tech, and AI in particular to significantly reduce harm in the short term by retro-fitting smart safety technology to road networks and existing fleets. This challenge should be led by technologists and government outside the vehicle manufacturing sector as manufacturers’ goal of maximising new vehicle sales aligns poorly with delivering social good from optimal road safety technology.
Solution: SmartSafe Network Technology
In this article, we will present an approach for using current technologies to build intelligent road safety networks. We call this SmartSafe Network technology.
SmartSafe works at the infrastructure level and operates with the existing vehicle fleet, delivering a leap forward in road safety at a cost of a few hundred dollars per vehicle, within a 5-year time frame. This is feasible with existing technologies, affordable, and politically tenable.
Why do we need to do something different?
Improved road safety has a direct impact on saving lives. In the New Zealand context (reflected in many OECD countries), the road toll has remained fairly static despite many campaigns around driver behaviour. These are expensive and slow to shift the needle because the major underlying factor is human fallibility.
Anecdotally, just about everyone admits to having made a mistake that could have seriously hurt someone when driving so it’s not simply a “good” vs “bad” driver issue. “Vision Zero” road safety strategies acknowledge human factors but planning is very evolutionary in nature (best practice = “do more of the same”) and doesn’t really explore the potential for revolutionary new technologies to shortcut the process.
The social cost of this is huge, estimated in New Zealand (with a population of 5 million) as $4.8 billion in 2017 across 378 fatalities and 3,000 serious injuries. Many OECD countries have extensive networks of high-risk roads and older vehicle fleets, making a strong case for retrofitted AI-powered safety infrastructure when compared against the costs and time-frames for replacing physical infrastructure or the entire vehicle fleets.
What is the current intelligent transport landscape?
Smart roads and safety systems fall under the umbrella of Intelligent Transport Systems (ITS), encompassing emerging tech as well as now-standard technologies such as automated speed cameras, road tolling, dynamic signage and traffic management. AI is used for a variety of tasks such as managing traffic density, and monitoring intersections for “near misses”.
Several Connected Vehicle pilots are operational which draw on existing vehicle-to-everything (V2x) communication standards and have the potential to form the basis for complete safety-oriented communication protocols.
There is a broad range of active R and D projects and commercial products on offer, highlighting that the technological “building blocks” required to underpin an intelligent road safety network have moved out of the lab and are commercially mature in most cases.
➜ Ancillary technologies such as 5G networks are being used to test “vehicle to everything” (V2X) communication and Bluetooth 5.0-based Bicycle to Vehicle (B2V) protocols are being explored to encompass vulnerable road users.
➜ Powerful AI-based computer vision models can now run on inexpensive, low power hardware.
➜ Trimble’s RTX technology provides centimetre accurate GPS positioning.
So what are the gaps?
Recent paradigm shifts in AI technology, enabling computer vision models to run on inexpensive, low-power hardware are yet to be exploited as an enabling technology for wide-area road safety networks.
“Smart road” technology projects tend to be fragmented across small “demonstrators” which don’t generate sufficient data to build a clear-cut case for retro-fitting safety technology at scale to existing vehicle fleets and road networks.
While pilot systems have been deployed to small numbers of test vehicles, there hasn’t been an effort to build a comprehensive, safety-focused system at a scale that can benefit from network effects and generate enough hard data to justify significant investment. In fairness, this is hard as it requires bringing multiple parties together – government, transport agencies, technologists, funders, as well as end-users.
How to fill these gaps?
Applying a “lean start-up” mindset will help to bridge these gaps. Imagine an angel investor has just provided $5M of seed funding with the goal of reducing the roll toll in New Zealand by 30% in 5 years. What would happen next?
A typical process would identify an initial group of users (“customers”) for whom you can solve a compelling problem and generate “traction” to attract further investment. This group could be “frequent drivers on a dangerous rural commuter route”. These customers would largely reside in the same community with road safety already an area of significant concern and discussion.
A prototype built around a single, well-trafficked route could achieve both community “buy-in” and a high degree of coverage for those users by focusing on their daily commuter route. Given time, the system can collect hard data on risk factors, accident rates, “near misses”, and the impact of safety interventions. Strong data and a reduction in accident rates will support further investment rounds and the enterprise can grow and reduce road accident harm.
Why should we use AI for road safety?
On a fundamental level, a distributed road safety network needs to know the location and velocity of moving objects on the road, as well as the topology of the road itself. If the network has this information on a continuous basis it can then assess risks and apply interventions to reduce the potential for an accident.
The latest generation of AI devices can derive velocity and position data solely from cameras without requiring Light Detection and Ranging (LIDAR) or other expensive sensors. If that data is communicated to an in-vehicle user interface, risks can be dynamically assessed and drivers provided with advance warning to avoid accidents before they happen.
Earlier computer vision techniques struggled with variations in lighting, orientation, and background but current algorithms are vastly better and would enable “plug and play” nodes to be added to a network where required, without any complex tuning.
While vision-based systems may not work as well in heavy fog or driving snow when compared to other technologies, a supplementary safety network doesn’t need to be active everywhere, all the time to deliver its core benefits – 90% or higher “uptime” is achievable and sufficient.
Adaptive Safe Path technology and digital convergence
The SmartSafe network has a novel “Adaptive Safe Path” feature that learns what constitutes a “safe” driving envelope on any given stretch of road by observing how vehicles have recently navigated it, with the assumption that most vehicles, most of the time, are driving safely for the conditions.
As an example, imagine a 400m stretch of windy road in rainy conditions. The system will create a model of a safe driving “line” for the conditions based on the median approach (position and velocity) taken by recent vehicles. Should a new vehicle deviate significantly from that line (either through excessive speed or road position) then risk indicators may be triggered. This approach enables the network to adapt to road conditions, road hazards, intersections or road works fairly quickly without depending on a central database for road topology.
AI can also be used by in-vehicle modules to detect a sleepy or distracted driver or erratic driving, both major causes of accidents. Again this is no longer a major technical stretch and can run on cheap hardware, providing older vehicles many of the key safety benefits of newer high-end vehicles.
Adaptive Safe Path enables a network to adapt to road conditions and hazards without depending on a central database for road topology.
Other technologies are also converging to support the case for intelligent safety networks, including IoT, solar, battery and communications tech, and improved GPS systems. 5G tech specifically supports this type of application, including “device to device” modes enabling vehicles and roadside infrastructure to communicate without the need for cellular coverage (Cellular-V2X (C-V2X)). The majority of supporting technologies are mature enough that implementation and deployment risks are well understood.
Modern, inexpensive AI hardware can be trained to recognise and track any type of object. This example built by Intranel’s AI team is running on $150 worth of hardware at >50 frames per second. Training the model to recognise different scooter brands only took a few hours.
II. The wait for autonomous vehicles
In talking about a road safety network it does raise the question of why we need smarter roads if we can just wait for smarter cars?
If “smarter cars” mean fully autonomous vehicles, the wait might be 20-30 years in most places before the technology becomes ubiquitous and the existing fleet is replaced. That equates to a lot of (preventable) harm through road accidents.
A key difference between fully autonomous vehicles and a supplementary road safety network like SmartSafe is that the latter doesn’t need to be right all the time to deliver on its core purpose – to warn drivers of potential hazards. In fact, it may even be better to not be “always-on” to prevent over-reliance or complacency when driving. Reliable autonomous vehicles are orders of magnitude more difficult to achieve than a supplementary safety system as the acceptable failure rate is close to zero. Most people won’t embrace computers that occasionally glitch and kill people.
“Smarter cars” also encompasses newer vehicles equipped with advanced safety features, including automated braking and collision warning and avoidance systems and Advanced driver-assistance systems (ADAS) that can take over routine driving tasks in some highway situations. Research indicates that Forward Collision Avoidance technology (FCAT) systems may prevent 30-50% of injury crashes, although publicly available data is scarce. This strengthens the argument for retrofitting some of these features into existing networks and vehicles.
A “Smarter roads” approach enables retrofitting the existing fleet and fills the gap between current technology and the ubiquity of autonomous vehicles.
Why build a network into road infrastructure?
There are fundamental performance and adoption benefits in integrating tracking into the road network itself and exposing public API interfaces that inexpensive equipment can access.
Proprietary in-vehicle systems are currently limited by what their own sensors can detect, and can’t exchange data with other vehicles or a network to receive comprehensive information about moving objects in their vicinity. While pure “vehicle to vehicle” (V2V) plays with high accuracy GPS, equipped vehicles can mitigate the need for roadside nodes. They do require a very high percentage of drivers to opt-in in order to be effective, and covering vulnerable road users like cyclists and pedestrians becomes much harder.
An integrated tracking approach with roadside nodes means all vehicles can be tracked and all road users can receive basic visible warning signals from roadside nodes, regardless of whether they have an in-vehicle module. Early adopters with vehicle modules in coverage areas receive an even better experience from a system that can track everything without gaps from non-participating vehicles.
As an example, a network-based approach would enable all approaching vehicles to be warned of an oncoming reckless driver, even if that driver occupies a non-connected vehicle without an in-vehicle module.
Retrofitting a single dashboard device to an existing vehicle that can communicate with a SmartSafe road safety network is much less expensive than upgrading an existing vehicle with a comprehensive sensor suite. Autonomous vehicles and other driver assistance tools work better when supported by an integrated road network.
Smart road infrastructure has broad benefits outside the SmartSafe network, including accelerating the safe deployment of autonomous vehicles. Tracking everything on the road provides key additional data to self-driving vehicles, boosting both safety and reliability. Open protocols for vehicles to access road networks and share data will significantly enhance the performance of any self-driving technology.
The role of government and manufacturers
While the NZ government is encouraging ITS (Intelligence Transport Systems) innovation, current transport agency spending is very much on physical infrastructure and conventional “proven” technology, with the development of high-tech safety systems primarily left to vehicle manufacturers.
Manufacturers ultimately want to maximise new vehicle sales, a goal which is poorly aligned with safety infrastructure as a social good – i.e. retro-fitting safety tech to an existing fleet doesn’t sell new vehicles.
Government, as the major funder of infrastructure, should play a role in collaborating on high-tech road safety systems operating at the infrastructure level, and supporting the development and adoption of protocols and standards around this. The potential for huge reductions in harm from leveraging existing technology means smart road safety infrastructure should be a strategic focus for public sector funding, alongside other ITS priorities. The private sector cannot be relied on to fill this gap due to current incentives focusing investment primarily on new vehicles. The government also has a key role in ensuring that data generated by new technologies is used appropriately, in line with local privacy protections and expectations.
III. An AI powered solution
SmartSafe – Concept design
The core of the SmartSafe concept network is based around roadside tracking nodes that know their location in relation to the road and can collaboratively track the location and velocity of vehicles, pedestrians, and cyclists in visual range. The tech within the roadside node is familiar, consisting of a digital camera, AI engine, GPS module, wireless connectivity (DSRC or similar protocol over 5G), and two redundant powered Ethernet connections to different base stations. Each node includes a high brightness display panel to provide basic visible alerts to drivers without in-vehicle modules.
Roadside nodes work together to share tracking vectors for individual vehicles to get better location and velocity estimates and minimise the effects of occlusion from other vehicles.
Base stations are spaced every 1km and power a 500m string of 10 nodes in either direction. They can draw power from the grid or a solar/battery combination. Cloud connectivity is through 4G/5G, a LoRa IoT connection, trunked radio, or satellite internet. That connection can be low bandwidth given AI functions occur in local hardware.
In-vehicle modules receive data from the road network on vehicle trajectories and road topology and calculate potential risks to drivers. If these exceed a threshold, there are several tiers of audio and visual alerts depending on the risk level. The vehicle module can act as a “black box” recorder with both accelerometers and road and driver-facing cameras. It can also use AI to generate alerts around driver fatigue or distraction and feed data back to the network if a driver or vehicle represents a hazard to other road users.
Nodes are modular and redundant and will self-adapt to their location, enabling a roading crew to drop in a replacement quickly and easily.
At the network level, the SmartSafe system uses 5G wireless V2i channels to provide an augmented driving service to which any equipped vehicle module (or roadside node) can subscribe via API. Real-time safety-oriented intelligence is delivered via edge compute AI devices embedded within the nodes rather than a centralised service in order to enhance privacy, reliability, and low latency.
This can be visualised as an edge compute “bubble” where the network within a 1km radius of a vehicle delivers velocity and position of other road users as well as data on road topology and potential hazards. The “bubble” can operate independently of cloud services or any other infrastructure, greatly enhancing overall system reliability and privacy.
How AI prevents accidents
Assuming a road safety network can track vehicles reliably, the question becomes how can this be used to reduce accidents?
For retrofitted systems, it isn’t practical for a system to take control of a vehicle, so the AI’s job is to give drivers advance warnings of hazards, whether they originate from the driver’s behaviour or another road user.
If we look at common accident factors on the right of Table 1 below and match them with detectable events on the left of the table we can see intuitively that advance warning will avoid or mitigate accidents caused by these factors.
Evidence from existing (and more basic) in-vehicle advanced safety systems indicates this conclusion is sound, but a lack of publicly available data to date makes quantifying benefits accurately very difficult without building a comprehensive prototype. The crash factors on the right make up a key causative factor in the vast majority of serious accidents – this is data that is publicly available in many countries, including through NZTA’s Crash Analysis System.
Heuristics around alerts can be fairly simple – for instance, if a driver pulls out to overtake and there is an oncoming vehicle at a dangerous closing velocity, audio, and visual alerts will be issued. Metrics around what constitutes “dangerous driving” are more complex and might involve a combination of weather conditions, speed and acceleration data, and crossing the centre line as well as whether a vehicle deviates substantially from the “Adaptive Safe Path” generated from previous vehicles. AI predictive models can improve over time by feeding back data from accidents or “near misses”. A network-based approach has the advantage that safety messages can be micro-targeted to a particular area – for instance, if an accident has blocked a road or if police nearby have initiated a pursuit of a dangerous driver.
Any safety warning system comes with behavioural challenges. If users feel overwhelmed with unnecessary warnings, they get sick of them and pull the plug.
AI has a major advantage in being able to recognise and adapt to different drivers and prioritise risks. As an example, if the AI learns a driver typically approaches an intersection at a fast rate before braking, it will adjust when it issues a warning appropriately. This learning mechanism helps ensure that warnings will be less likely to be ignored.
For some drivers, there is a risk that the absence of a warning could be taken as a green light to initiate a dangerous maneuver, e.g. overtaking on a blind corner if an “oncoming vehicle” alert isn’t triggered when they pull out. This can be mitigated by system design, e.g. recognising the maneuver as dangerous due to road topology and issuing the same warning as for an oncoming vehicle (or any other major imminent hazard). Keeping drivers aware that the safety system may not always be present or “always-on”, may minimise over-reliance or complacency.
Interestingly, work by SAE indicates that existing in-vehicle Advanced Driver-Assistance Systems (ADAS) are (so far) not encouraging driver inattentiveness. Being accountable for reckless driving identified by the system would also go some way to encouraging better practices but also has privacy implications and risks reducing uptake – drivers will not want to be “spied” on and any non-anonymised reporting of behaviour will have to be negotiated with communities, reviewed by a human, and reserved for the most serious cases.
The first thing many New Zealand drivers do on acquiring an imported vehicle is disable the automated speed warning that beeps over 110km/h.
The value of data
A huge benefit of a road safety tracking network is that, even on a small scale (less than a few thousand vehicles), it yields comprehensive data not just on the lead up to accidents but also “near misses”, which currently don’t get recorded and probably outnumber accidents 10 to 1.
Driving data that precedes accidents or near misses can be captured to continuously train better AI and risk models and enable better targeting of risky behaviours. This targeting can be much more nuanced than say instantaneous speed cameras and can have an educative focus, with personalised feedback available to drivers. At the policy level, strong data enables better safety interventions and informs transport policy and strategy development.
AI-Powered road safety networks like SmartSafe generate comprehensive data to build and improve risk models.
Privacy concerns are legitimate with any network that tracks movements of vehicles or people. Opinions will vary, ranging from an expectation of almost complete privacy to the view that driving is a privilege and drivers don’t necessarily have a right for their behaviour to be outside the purview of the state. The status quo sits most closely with the latter stance, with tacit acceptance of broad deployment of enforcement measures and traffic cameras. Within developed urban centres, licence plate recognition technology capable of tracking an individual vehicle around a city is already common.
Privacy protections lie not with the absence of such technologies (they already exist) but with legal protections restricting how data is gathered and used, and transparency and oversight around this. Privacy can be managed with system design, especially where AI happens on local nodes and video data isn’t routinely sent to a central server.
From a privacy risk perspective, there’s a big difference between hoovering up every licence plate and sending it to a central server versus local roadside nodes receiving a list of stolen vehicle plates to look out for. For some drivers, enhanced monitoring by insurance companies, parents, or employers may be appropriate as part of the conditions to access a vehicle.
Centralising non-anomymised data can be triggered only in the event of a serious accident, as would accessing in-vehicle module data. This is similar to the situation now with some drivers installing cameras for their own peace of mind.
Politically, significant public backlash is less likely if initiatives are community-led, and the focus on safety and saving lives is strictly maintained. The proposed system is designed to protect participating drivers, without uptake needing to be universal in pilot areas – effectively an “opt-in” system. The initial appeal is likely to lie with parents of younger drivers and regular commuters on dangerous routes, with the insurance industry another clear beneficiary.
The threshold for non-anonymised reporting of risky or illegal driving behaviour should be negotiated with communities to ensure “buy-in” is present, and human review applied to reports, prior to sanctions being applied.
Costs and return on investment
Costs will vary between regions and geographies but modelling in the New Zealand context indicates an ROI over a 20-year lifespan of ~700%, meaning for every dollar spent on the network $7 in measured “social harm reduction” will be generated through reduced injuries and fatalities. Installation is assumed to be primarily on state highways which are responsible for a higher portion of harm per kilometre and present less complex environments for monitoring. Social costs of accidents and injuries use New Zealand Transport Agency figures.
Right now new technology development in the transport space is focused on a space race to build autonomous vehicles. This is laudable, but the 20-30 year time frame to fully replace existing fleets means there’s a 15 to a 20-year gap where road accident harm will continue at similar levels to now.
This gap can be addressed in the short-term by retro-fitting intelligent road safety systems to existing infrastructure, built around AI and other proven technologies. Systems like the proposed SmartSafe network concept will generate raw data to enable safety interventions as well as supporting a myriad of in-vehicle systems, including accelerating the roll-out of autonomous vehicles.
The challenge is about reducing social harm through optimal use of high tech public infrastructure and as such government and transport agencies should play a major role. Vehicle manufacturers are valuable stakeholders but their focus on new vehicle sales means they cannot be relied on to lead the delivery of optimal solutions.
A smart safety network can be built quickly relative to autonomous vehicle tech in that it doesn’t attempt to replace the driver, but rather has a supplementary role in providing advance warning prior to a risk manifesting as an accident. This delivers major benefits without needing to be 100% accurate at all times and in all conditions.
A razor-sharp focus on safety and application of lean business principles to building a beta system at a modest scale will generate measurable harm reduction and a valuable data set to build risk models. This will inform data-driven policy around interventions and help train AI-powered “driver assistants” to mitigate risk prior to an accident occurring. Hard data and an “open IP” approach will facilitate subsequent investment and roll-out, with the potential to meaningfully reduce harm at the national level within 5 years.
Geographies like New Zealand with a large network of high-risk highways, high accident rates, and older vehicle fleets are especially well-placed to benefit from this technology.
While there are known risks in managing human factors and privacy concerns, the challenge is an engineering one and not “blue-sky” research. Lives can be saved with existing technologies in a way that is both affordable and politically tenable.
About the author
Adam Lyness, Founder and Business Development Manager at Intranel Consulting Services. Adam is an entrepreneur and an emerging technology leader. Through his role with Intranel Consulting, he enables business partners to accelerate technology delivery using a lean, business-driven approach.