An automatic driving car is likely to cause a collision accident if the pedestrian's skin color is dark
by ke wen
It is already known that some face recognition systems are struggling to accurately identify people with dark skin color, but the artificial intelligence (AI) system newly installed in the automatic driving car also It is clear that we have similar problems. According to research, it seems that the probability that a pedestrian with dark skin color encounters an accident with an automatic driving car will increase.
Predictive Inequity in Object Detection
(PDF) https://arxiv.org/pdf/1902.11097.pdf
Self-Driving Cars May Hit People With Darker Skin More Often
https://futurism.com/the-byte/biased-self-driving-cars-darker-skin
Researchers at the Georgia Institute of Technology have investigated the accuracy of the eight AI models used in the state-of-the-art object recognition system used for automatic driving cars, etc., and posted the results in arXiv of the paper submission site as a paper doing. The object recognition system investigated by the research team is used by the automatic driving car to correctly recognize road signs, pedestrians, and other objects.
In the survey, based on the score of " Fitzpatrick classification " which is generally used for classifying the color of human skin, pedestrian photographs to be checked by the object recognition system are classified as "a category with a weak skin color" "Category of deep color of skin" category. Then, in the object recognition system, a pedestrian carries out a test of letting a pedestrian judge which one of "a person with a weak skin color", "a person with a dark skin color", "a skin color can not be identified but a human being" or "not a person" doing.
As a result of the test, the object recognition system shows "uniformly inferior performance" for pedestrians with "light brown", "brown", "black" which are said to have dark skin color in Fitzpatrick classification It is clear that it is. According to researchers, the group including pedestrian images with darker skin color reduces the accuracy of the object recognition system by 5% on average, even though considering variables related to the timing of taking pictures such as daytime or nighttime It is said that.
by Tomo Nogi
According to the research result of the research team of Georgia Institute of Technology, if the object recognition system of the automatic driving car is spread with the accuracy as it is, the probability of encountering a traffic accident will change according to the color of the skin. However, based on the findings, the research group says "By using more images of pedestrians with darker skin color in the data set, we will be able to detect pedestrians more accurately."
Related Posts: