Georgia Institute of Technology Finds Autonomous Cars Struggle to Detect Pedestrians With Dark Skin
There are a lot of concerns in the air when it comes to autonomous vehicles and safety. One of the more existential questions that's lingering in the air is the moral dilemma of whether autonomous vehicles will prioritize the lives of the passengers inside the vehicle or those of pedestrians outside the automobile. Then, there's the issue of how they'll handle in the real world alongside human drivers.
Now, there's another issue to worry about. According to a study conducted by the Georgia Institute of Technology, systems found in autonomous vehicles have trouble detecting pedestrians with dark skin.
Pedestrians With Dark Skin Could Be At Risk
Researchers from the institute started out with a simple enough goal, aiming to find how object detection systems in state-of-the-art vehicles performed with detecting pedestrians with different shades of skin. The researchers utilized the Fitzpatrick scale, which is a system that classifies skin tones into six different categories. The classification system was originally based upon how six different skin tones respond to sun exposure. A low ranking denotes a lighter skin tone, while a higher ranking relates to a darker skin tone.
Using the Fitzpatrick scale, the researchers separated photos of individuals into two groups. One group was made up of pedestrians that were classified on the lighter shade of the scale, while the second group were part of the scale's darker shade of skin tones. The findings weren't very promising.
The Georgia Institute of Technology's study found that object-detection systems were 5 percent less accurate when it came to detecting individuals that were in the group that had darker skin tones according to the Fitzpatrick scale. Variables like obstructed views or the time of day didn't help boost those numbers, either.
"The main takeaway from our work is that vision systems that share common structures to the ones we tested should be looked at more closely," Jamie Morgenstern, one of study's authors, told Vox.
Why The Discrepancy?
While the study's findings are alarming and eye opening, it's certainly not perfect. As Vox points out, researchers weren't able to use object-detection models that are currently being used in driverless vehicles. The outlet also claims that the study didn't "leverage any training datasets" that are currently being utilized in self-driving cars. Instead, what the researchers did do was test models that academic researchers use.
So what the study actually shows us, is how algorithmic bias works, and how it could possibly stop autonomous vehicles from being 100 percent foolproof on the road. Object-detection systems, as Vox outlines, have had multiple issues in the past, including when Google's image-recognition system labeled photos of African Americans as "gorillas" back in 2015.
The issue with algorithmic systems comes from the examples they're learning from, claims Vox. The researchers from Georgia Institute of Technology mimics the outlet's thinking, claiming that object-detection models that were tested in the study were mostly being trained on pedestrians with light skin. The models also didn't put enough "weight" on learning from when a pedestrian with dark skin was detected.
While there are multiple things that companies could do moving forward to ensure that systems on autonomous vehicles detect pedestrians of all skin tones, the study proposes that companies pay more attention to their training methods.
- Volkswagen Announces EV Plants in China to Compete With Tesla
- EVs Expected to Account For Half of New-Car Sales by 2040
- Germany Opens First Electric Highway to Charge Hybrid Semi-Trucks
- Illinois Becomes Latest State to Charge EV Owners With Annual Fee
- May Mobility to Begin Testing Autonomous Shuttles in Rhode Island
- General Motors Holding Talks to Sell Lordstown Factory to Startup Workhorse
- Honda and Toyota Scrambling to Find Funds to Develop EVs, New Tech
- Tesla Claims Affordable Autonomous Cars Aren’t Possible Without China