In the recently published “Annual Report on Road Accidents in India-2022” by the Ministry of Road Transport and Highways, the statistics paint a concerning picture: 4,61,312 road accidents were reported across States and Union Territories, resulting in 1,68,491 fatalities and 4,43,366 injuries. These figures reflect a significant increase compared to the previous year, with accidents up by 11.9%, fatalities by 9.4%, and injuries by 15.3%. The primary causes of these accidents include over-speeding, rash driving, violation of traffic rules, failure to understand signs, fatigue, and driving under the influence of alcohol. In this context, Vinay Rai, Executive Vice President of Engineering at Netradyne, recently spoke to Tech Achieve Media to discuss the pressing issue of road safety and how advanced technologies are playing a crucial role in mitigating these risks. With a wealth of experience in developing AI-driven solutions for road safety, Vinay shares insights on the impact of these technologies and the future of safer driving in India.
TAM: With road safety being a critical issue worldwide, particularly in countries like India, how can deep technology solutions effectively address behavioral factors such as drowsiness and distracted driving? What role do real-time monitoring and behavioral analytics play in influencing and improving driver behavior?
Vinay Rai: These statistics are indeed troubling, and something we’re very aware of. However, the good news is that technology today offers solutions to address these behavioral challenges. Interestingly, some of the technology used to tackle this issue is similar to what’s being developed for autonomous vehicles. For example, with an autonomous vehicle, you’re not focusing on the driver’s behavior—it’s designed to drive itself. But the underlying infrastructure or the information the autonomous vehicle relies on is quite similar to what can be used here.
Consider a traffic light: an autonomous vehicle is programmed to slow down and stop when approaching one. Similarly, for a human driver, we can develop systems that detect the presence of a traffic light and determine if the driver is responding correctly—whether they slow down or stop as they should.
When we talk about behavioral aspects, it’s like analyzing how a driver responds to traffic lights over time. If a driver encounters hundreds or thousands of traffic lights in a month, do they consistently obey them, or do they tend to ignore them? That’s the kind of behavior we can now monitor and assess.
Netradyne offers a solution that can be implemented in any vehicle today—you don’t have to wait for fully autonomous cars. Even though autonomy has different levels, which we won’t delve into now, technologies like IoT, deep learning, and cloud computing, when integrated, allow us to determine whether a driver is consistently driving safely—not just as a general habit, but in specific moments as well. Here’s how it works: An IoT device, much like a smartphone, uses advanced computation and deep learning models. It can see what you see and process that information, both inside and outside the vehicle. This allows it to provide real-time feedback to the driver and send data to the cloud. With this technology, we can identify which drivers are habitually safe and which ones may pose a risk.
Because this system is cloud-connected, the data can be reviewed later by the driver or their manager. In fleet situations, fleet managers can show drivers specific incidents, giving them a chance to reflect on their actions. Often, we’re not aware of our mistakes, especially if they’ve become habitual. But when an unbiased system shows us proof of these mistakes, it prompts reflection. Most of us want to do the right thing, and with this feedback, we can improve. We’ve observed this with many of our deployments—our customers have reported a decrease in the number of alerts and violations, whether voluntary or involuntary, over time.
TAM: In the context of deep technology applications in transportation, how are AI, edge computing, and computer vision redefining the landscape of fleet safety management? What are the most significant advancements that these technologies bring to the table in terms of preventing accidents and ensuring driver safety?
Vinay Rai: The most important point is that we don’t need to wait for autonomous vehicles to hit the roads to improve safety. Even when autonomous vehicles do become common, they won’t solve the problem entirely unless every vehicle is autonomous—and even then, they’ll need to be trained specifically for the Indian context, which will take time. However, the technology we already have can help prevent accidents right now.
We’ve been offering this solution for nine years, and with increasing awareness of road safety in India, it’s more relevant than ever. The technology works by providing instant feedback if you’re distracted, drowsy, or tired. It’s like having a friend in the car who alerts you when you’re about to make a mistake—whether you’re looking at your phone or engaging in risky driving behavior. This real-time feedback can prevent accidents as they happen. But that’s just one aspect. The other crucial element ties into the behavioral aspect. Many of us occasionally glance at our phones while driving, thinking nothing will happen. However, if this behavior becomes a habit, it’s only a matter of time before it leads to an accident—whether it’s next week, next month, or next year. That’s why behavioral training and coaching are so important. We don’t just provide instant feedback to correct behavior; we also collect data over time to give drivers a safety score.
For example, if you drive 200 hours in a month and encounter 500 traffic lights, how often do you follow the rules, avoid speeding, or resist checking your phone? Each instance might not immediately cause an accident, but over time, these behaviors increase the risk. Today’s technology allows us to address this on two levels: providing instant corrections and accumulating data to offer long-term coaching, helping drivers avoid risky situations altogether.
TAM: How do you address privacy concerns and ensure the ethical use of AI-driven monitoring systems? What measures are in place to protect drivers’ data and maintain transparency with clients and end-users?
Vinay Rai: First, we are a B2B company, though there are also B2C solutions. While there are some minor differences, most practices remain the same. At the core, we ensure that access to data is restricted to only those who need it. Who needs access? One key group is the drivers themselves—they should be able to see all relevant data about their own driving if they wish.
Secondly, our system processes and records everything it detects on the device, but not all driving data is uploaded to the cloud. Only incidents involving risky driving are recorded and uploaded because that information is crucial for driver safety and coaching. For instance, if you drive 200 hours in a month and have just one violation—like running a red light—only that specific one-minute video is captured, while the rest of the data is overwritten over time. We do keep some data temporarily on the device, in case it’s needed later for incidents like road rage, or if a fleet wants to review recent data. However, after about a month, the data is erased.
Additionally, we give our customers the option to blur the video. For example, if you’re driving with a passenger, the system can blur the passenger’s identity, especially in sensitive cases like taxis or cabs. We take privacy very seriously and offer this blurring feature so that only necessary details are visible. Some customers might prefer not to blur the video to capture more information, but others focus solely on how the driver is performing, in which case we blur everything else. In many cases, we also blur the driver’s face unless specific details are needed, such as when detecting if the driver is drowsy or distracted.
It’s not just about the driver’s privacy, though. We also protect the privacy of others, like pedestrians or people’s homes that the vehicle passes by. For instance, if a vehicle records video while driving past your house, we blur any identifying features to protect your privacy.
Regarding data, it’s crucial to strike a balance. Without data, we can’t train our models. Think of it like a child—learning is essential, but it shouldn’t come at the cost of others’ privacy. Our privacy policy ensures that while we train our systems, we don’t track personal details like who was driving. We only need to know that someone was driving and encountered certain scenarios, such as a traffic light. We ensure that no personal details are stored long-term for training purposes.
We’ve gone even further with our privacy measures. For example, even knowing the geolocation data could inadvertently reveal a driver’s identity if it shows repeated activity near a specific home. To prevent this, we use additional algorithms that anonymize such data, ensuring that privacy is maintained while still collecting necessary information for training.
Training models like those for generative AI requires vast amounts of data—our systems have logged 15 billion miles of data, which is equivalent to 2,500 crore kilometers. This extensive data is crucial for training generative AI models. However, for our specific model training—like detecting pedestrians, vehicles, traffic lights, or license plates—we require much smaller datasets. We clean this data, apply privacy filters, and store only what’s necessary for training.
We adhere to the concept of “privacy by design,” and we have a VP of Privacy who acts as an internal auditor, ensuring that our engineering teams comply with the latest global standards. Many of the measures I’ve described are a direct result of these rigorous checks and balances.
TAM: Can you provide examples of how Netradyne’s technology has demonstrated effectiveness across different sectors, such as oil and gas, trucking and logistics, and passenger transport? Could you give us a few examples ?
Vinay Rai: The effectiveness of our solution is consistent across various industries. Many of our customers have reported a reduction in accident rates between 50% to 70% after deploying our solution. For instance, some of our larger clients initially discovered, through our data, that their drivers were frequently engaging in what they assumed to be safe driving behaviors, but which were actually quite risky. Within just three months, one particular type of alert decreased by 80% for a specific customer. This trend is evident across industries, whether it’s long-haul trucking, last-mile delivery, or school bus transportation.
Part of this improvement comes from drivers knowing they are being observed, which naturally encourages safer behavior. Additionally, it’s crucial to educate drivers that the system is for their safety, not for surveillance. We emphasize that our technology is not a surveillance tool because nobody likes being surveilled. Everyone wants to do the right thing, and our system helps them achieve that. Across all industries, we’ve seen a consistent reduction in accidents and in the number of alerts, which are often early indicators of potential accidents.
In the oil and gas industry in India, for example, some customers have mandated that if our device is not functioning—whether due to a hardware issue or deliberate tampering—the vehicle should not be driven. This reflects the level of trust in our system. Some of our customers even have 24/7 operations teams monitoring the alerts. If the system detects that a driver is drowsy, they are notified immediately, and they call the driver to ensure their safety. In one such instance, a customer thanked us for the system after they were able to stop a drowsy driver before a serious accident occurred.
A specific example from India is Writer Safeguard, a company that operates yellow vans carrying cash to replenish ATMs. This company, part of the Hitachi Group, deployed our technology and tracked various KPIs related to the causes of accidents, such as speeding, drowsiness, and distractions. Within three months, they saw a significant reduction in these KPIs, and overall, they experienced a 50% reduction in accidents within nine months. This information was publicly shared, and we can provide more details if needed. There are many other examples, though we are not always allowed to name our clients. Writer Safeguard’s experience is particularly noteworthy because their vehicles operate not only on highways but also in small towns and villages, where road conditions can be quite different from those in metro cities.
TAM: Looking to the future, what innovations or advancements can we expect from Netradyne in the realm of road safety technology
Vinay Rai: The future is always exciting. When we started back in 2015, our work was groundbreaking, but it also came with significant risk—there was uncertainty about whether the market would accept our solution. Fortunately, we’ve now reached a point where there’s clear acceptance.
Looking ahead, one area of growth is gaining more government support through policies. For example, if the government mandates that everyone must have this kind of solution, it would be a great step forward. This could also be tied to incentives, like reducing insurance premiums for those who use our solution. We already have a “green zone score,” which is like a CIBIL score but for safety. Just as a CIBIL score reflects creditworthiness, the green zone score reflects how safely someone drives. Many of our customers use this score to reward drivers with bonuses for maintaining a high score. Similarly, it could be used by insurance companies to lower premiums for safer drivers. Government support in this area would be invaluable.
On the technology front, there are many exciting developments. Generative AI is a hot topic these days. While some of the buzz is hype, there’s also real potential, and we aim to be on the side that solves real problems. We’re particularly excited about our 15 billion miles of data. In AI, the cleanliness of data is crucial. The cleaner the data, the more accurate the model. It’s like giving a child clear study materials versus confusing them with extra, irrelevant information. With our vast, noise-free data, we’re in an excellent position to leverage this for various applications, including the development of autonomous vehicles.
Autonomous vehicle development has been ongoing for about 15 years, and while there’s been progress, there’s still a lot to be done. One challenge is knowing if all possible scenarios have been covered. Human-designed test cases can cover a lot, but there’s always a “long tail” of rare situations that only real-world data can reveal. Most companies working on autonomous vehicles haven’t driven nearly as much as we have; many boast about 1,000 or 5,000 kilometers of safe driving, but that’s not enough to reflect the true complexity of the world. With our billions of miles of data, we’ve likely covered that “long tail” of edge cases.
One use of our data is to train autonomous vehicles, but even more exciting is the potential to anticipate what will happen on the road. Large language models (LLMs) work by predicting the next most logical word in a sequence. Similarly, with our vast data, we have the ability to predict what’s likely to happen next on the road. This can be invaluable not just for building intelligence into autonomous vehicles but also for providing anticipatory feedback to human drivers. Imagine being able to warn a driver about potential hazards based on subtle cues, like something protruding into their peripheral vision—whether it’s a bus, a dog, or another obstacle. This is the kind of predictive capability we’re working on in our labs. While it’s not yet a finished solution, it’s an exciting area of development.
