Study: AVs may struggle to detect pedestrians with dark skin
- A study from Georgia Tech University warns that some technology used in autonomous vehicles (AVs) has trouble detecting pedestrians with darker skin.
- The report, entitled “Predictive Equity in Object Detection,” analyzed AV footage from New York City, San Francisco, Berkeley, CA and San Jose, CA, and studied how eight image recognition systems reacted to pedestrians’ different skin tones, as measured on the Fitzpatrick skin type scale. Researchers found the systems have shown "uniformly poorer performance" when detecting those between 4 and 6 — darker skin tones — on the Fitzpatrick scale.
- Even when researchers accounted for changes in weather or the amount of daylight, they found the models were 5% less accurate when detecting pedestrians with darker skin tones.
This latest research comes as facial recognition technology has been under increasing scrutiny. Amazon has been offering its software, called Rekognition, to police departments nationwide, but is the subject of “serious concerns” from lawmakers who worry about its impact on communities of color. The American Civil Liberties Union (ACLU) has also warned the technology is "primed for abuse in the hands of governments,” although cities like Orlando, FL have piloted the technology for uses including crowd control and tracking violent criminals. According to a public poll, few people support restricting the technology’s use.
News of this study’s conclusions will likely add to the safety concerns over AVs, which were amplified last year when a pedestrian was fatally struck in Tempe, AZ by an Uber-owned AV being tested on city streets. It comes with companies still refining the technology to help it make ethical decisions on the roads, and with cities experiment with updating infrastructure or installing new systems to help AVs communicate with each other and their surroundings.
Companies such as Waymo have continued to herald advances in AV technology, like the ability to obey hand signals from police officers, while LiDAR and other radar systems continue to evolve and be more responsive to sudden obstacles like pedestrians. But this study shows there is still a lot of work ahead for the companies testing and researching AVs to ensure they are truly safe.
Follow Chris Teale on Twitter