Cornell Technology Can Predict Car Accidents Before They Happen
Other than the occasional collision that result from an unexpected deer or nearly-invisible patch of black ice, car wrecks are almost entirely the result of human error. Exterior safety features like lane keep assist and automatic emergency braking are doing a good job of minimizing the consequences of those errors, but the driver still has to make a mistake before those systems can react. What if a car could know what its driver was going to do beforehand and could give a warning or react to prevent a mistake from happening in the first place? New technology being developed at Cornell and Stanford could do just that.
In the same way that body language reveals whether someone is bored, excited, sleepy, or romantically interested, body language can also give away what a driver is about to do. Whether a driver is about to change lanes, make a turn or accelerate from a stop, small movements of the head, neck, and body indicate that before the action is initiated. Combining that information with information from external sensors, this system is able to guess what the driver is going to do, monitor the road around the driver, and give a warning if what the driver is about to do may be dangerous.
Cornell’s Ashutosh Saxena developed the system with his colleagues by recording 10 drivers as they drove a collective 1,200 miles. Combining that video footage with footage from forward-facing cameras on the cars they were driving, they were able to gather data on the drivers’ body language. The program they created with results of that data analysis was then used on a different set of drivers and ended up being more than 75% accurate. It could also, on average, predict a turn or lane change 3.5 seconds before it happened.
It’s still in its early stages and isn’t a perfect system just yet, but Saxena is working hard to fix the kinks. Shadows, for example, can mess with the face detection software, as can the driver’s interactions with passengers. Adding eye tracking, sensors on the steering wheel, and sensors on the pedals would likely increase the system’s accuracy. Integrating GPS data could also tell drivers when turns they’re about to make are illegal.
While the concept is intriguing and could potentially help prevent wrecks, the biggest challenge Saxena and his team are going to face with bringing it to production is the fact that while their system helps make people better drivers, autonomous driving technology is working to remove the driver altogether. Production cars like the Mercedes-Benz S-Class are already nearly self-driving, and Tesla promises that the Autopilot feature the Model S will be receiving soon will allow a driver to go from Seattle to San Francisco without touching the steering wheel.
City driving is, of course, much more complicated than highway driving, and while the future of cars is almost certainly a driverless one, it’s going to be a while before truly autonomous vehicles are roaming the streets. Nearly autonomous cars will probably become fairly common in the next five years, but navigating a city after a blizzard is no job for Autopilot just yet. With that in mind, perhaps predictive, driver-monitoring systems will have a place in cars after all, helping drivers make safer choices in the times that they are in charge of what the car does.
Saxena is seeing some interest from automakers, so it looks like they see the potential of his system too. They may be investing heavily in technology that would make his system irrelevant, but they also have to all know that completely self-driving cars are still a long way from being viable, much less legal. Investing in a system that can predict driver behavior and stop crashes before they happen could give a company like Volvo a safety advantage in the marketplace. Even if it never makes it to market, it’s still a fascinating piece of technology.