Social distancing is one of the main weapons in the fight against the coronavirus 鈥� and computer scientists have used a database of public cameras to keep track of how well we are following the guidelines.
Since April, Isha Ghodgaonkar at Purdue University, Indiana, and her colleagues have gathered around 0.5 terabytes of data per week from 11,140 public cameras connected to the internet.
More than 10.4 million images from the webcams have been run through deep-learning neural networks that automatically detect objects and differentiate them from people. Bounding boxes are drawn around the people, and the neural networks can then calculate how far apart they are from one another 鈥� and so whether or not they are practising social distancing.
Advertisement
The researchers found that people are social distancing, to a degree. Crowd densities were lower after authorities imposed lockdowns and higher when restrictions were eased, with social distancing following that principle, too. In the UK, for instance, after the shift in public messaging from 鈥渟tay at home鈥� to 鈥渟tay alert鈥� on 10 May, the number of people captured by the cameras increased.
There are blind spots: Ghodgaonkar鈥檚 team managed to find just four public webcams in Russia, compared with more than 10,000 across the US. In addition, image-detection software assumes that all humans are standing upright, and are all the same height 鈥� 1.65 metres 鈥� in order to calculate their bounding boxes, depth in the image and so their distance. Deep learning also can鈥檛 determine whether people not observing social distancing are strangers or within the same household.
Governments are utilising location-tracking data from Google and Apple to observe movements. Monitoring cameras is a better method, says Ghodgaonkar. 鈥淟ocation tracking might be slightly biased because it鈥檚 dependent on people opting in,鈥� she says.
Others disagree. 鈥淪olutions based on aggregated and anonymised mobile phone data pose fewer [privacy] risks than solutions based on surveillance camera data,鈥� says Sam Gilbert at the University of Cambridge. He worries that benign computer vision could be used for nefarious means in politically oppressive countries.
George Thiruvathukal at Loyola University Chicago and a co-author of the paper says they 鈥渨anted to be as least invasive as possible and respect people鈥檚 privacy鈥�, pointing out that the camera images are publicly accessible.
Reference:
