By Arie Zilberman, CEO and Co-Founder of Keymakr
A whopping 42,915 people died in car crashes in 2021 in the U.S. alone, a stunning statistic considering how much tech is flooding the automobile market. Much of that tech is aimed at producing self-driving vehicles, which creates a new front in the effort to boost safety on the road.
Consumers are still skeptical that self-driving cars can actually be safe enough, which to some degree makes sense. Autonomous-vehicle development has faced more obstacles than originally anticipated in the past few years, mainly due to the lack of technology that actually works like it’s supposed to.
In early September, General Motors recalled Cruise, its autonomous driving software, after a recent crash during its test run, making the public even more skeptical of the efficacy of the emerging technology. Nevertheless, autonomous-driving technology company Mobileye went through a successful IPO just last month, signaling the demand for reliable autonomous vehicle infrastructure. With Mobileye, Tesla, and other big name brands investing a lot of resources in the sector, investors will want to see results to justify their growing market caps.
The need for better datasets
Building autonomous vehicles that can actually be driven on the road, and at scale, requires companies to create complex datasets that reflect real-world conditions, and train them to ensure passenger safety. That task differs from prominent AI applications today, such as personalized shopping, Siri-style voice assistants, or virtual travel agents, because it requires the heavy use of imaging and video to collect the data necessary to actually map out physical spaces and patterns in those spaces.
To collect that data, companies often mount logs to the outsides of cars, including lidar sensors, cameras, radar, and GPS. They leverage it to train machine learning models responsible for perception, prediction, and motion-planning capabilities. These systems make sense of the world and the objects within it and dictate the paths vehicles ultimately take.
As such, that data is heavily location-dependent. In England, for example, people drive on the opposite side of the road than in America. Even within the U.S., drivers in New York City are generally more aggressive than those in the suburbs. Self-driving vehicles need to be able to adapt to their environments, and machine learning models need to take this into account. That includes weather—rain, snow, hail, and fog.
Improving the driving experience
A General Motors autonomous vehicle failed back in June due to the software “incorrectly predicting another vehicle’s path,” according to CNBC, a challenge that only more precise video annotation can fix. Annotation tools are commonly used in developing AI applications in order to label objects of interest or segment pixels in target classes, in each frame of the video.
Video annotation can help AI study patterns and replicate the complex movements that occur on the road. Through the increased use of video annotation, not only will autonomous driving become more dependable, and therefore commercial, but roads will become safer for all.
To build safer roads, data annotation must eliminate biases. This can be done by gathering data from different regions and different countries where people drive differently and where different street signs, street lights, cars, and objects may factor in.
For advanced driver-assistance systems (ADAS) there is a need for datasets from different demographics: age, gender, ethnicity. Additionally data creation needs to be capable of mimicking different types of behavior behind the wheel, such as sleepiness, distractedness, or even a passenger in the backseat talking to the driver.
Errors in classification or the lack of quality data can lead to deadly consequences. The wrong classification of a line in the road can be the determining factor in a car crash, and many companies unfortunately aren’t equipped for this high-level of data annotation. These mistakes are often hard to catch and monitor due to the large quantity of data.
These types of mistakes tend to happen when the workforce is outsourced to external teams. When working with large projects and demanding delivery times, as often is the case with autonomous vehicles, there is a need to process data at a rapid pace, and for many companies this just isn’t possible. In this case outsourcing the role of data annotation to a company specializing in annotation may be the best option.
Nvidia, whose DRIVE Hyperion offers a complete development platform and reference architecture being leveraged by companies such Mercedes and Volvo, has mapped over 300,000 miles of roads across the globe to improve safety. Additionally, iMerit, whose team of data annotation specialists have annotated 250 million data points, leading the way in the development of smart sensors and high-definition maps. These companies and many other prominent and upstart players are making strides in the development of autonomous vehicle technology and manufacturing, enhancing safety through object recognition and behavior monitoring.
Object recognition and behavior-monitoring capabilities ensure AI-powered vehicles can navigate safely on dynamic and chaotic roads in real-world settings, paving the way for widespread acceptance of autonomous driving.
Object recognition doesn't only apply to autonomous driving, but to regular vehicles as well. AI models can be trained to recognize important objects, such as keys and mobile phones, or even people, like children in the backseat, and warn drivers and passengers when they are left in the vehicle. AI models can also be trained to interpret the behavior of drivers. If a driver falls asleep or is not paying attention, it can warn the driver to pull over. In order to achieve these targets and lead the way for more innovation to follow, the focus must be on improving datasets.
About the author:
Arie Zilberman is an entrepreneur who founded and currently runs several successful businesses dating back to 2000. He has vast amounts of experience in constructing, managing, and leading large teams. Zilberman founded Keymakr, a data-annotation service provider delivering high quality datasets to AI innovators, in 2015 and currently serves as CEO.
The views and opinions expressed herein are the views and opinions of the author and do not necessarily reflect those of Nasdaq, Inc.