Facial recognition can get better at spotting people from a distance, new paper says
Surveillance cameras with facial recognition are covering more of our cities. But the systems are facing a problem: they are relatively shortsighted.
A new study published in Nature looks at how to optimize surveillance cameras for matching face biometrics from a distance.
The facial recognition method is based on deep learning from a dataset that includes the distance of the individuals from the camera, the focal length of the image sensors and the size in pixels of the target face. The extended dataset was created using the Georgia Tech Face and Quality Dataset for Distance Faces.
The method was then tested on different surveillance camera image sensors, with several sensors managing to achieve an average accuracy above 99 percent in the recognition process. It also allowed the researchers to calculate the maximum distance for each sensor to attain the required accuracy for facial recognition.
The researchers claim that their method could be crucial in security applications in smart cities.
To make the best of the new datasets, the paper advises governments and enterprises introducing surveillance cameras to first assess image sensors by estimating their accuracy for facial recognition. This allows them to choose higher-accuracy image sensors for areas that are large and require surveillance cameras to recognize faces from a distance. Similarly, they can save costs by buying lower-accuracy image sensors for cameras covering smaller areas.
“The novelty of this work comes from providing a method for selecting an image sensor that effectively balances accuracy and cost for a specific distance,” the paper notes.
The study was published by a team of researchers from the University of Alicante in Spain, the SGH Warsaw School of Economics in Poland, and Effat University in Saudi Arabia.
The research has a few caveats. The results come from a synthetic database, with researchers recommending further tests using real video surveillance images that contain more variety.
The U.S. federal government has been funding research into long-distance biometric identification, including face biometrics.