New research on CCTV camera prevalence in cities uses Google street view
New research from Stanford University has estimated the numbers of CCTV cameras in cities across the world using machine vision and Google street view. Ron Alalouff reports.
Do you know how many CCTV cameras there are there in London, or in other major cities around the world? While there have been attempts to estimate these numbers, systematic surveys of camera density are hard to come by. While some studies have sought to estimate the number of CCTV cameras installed in a city, only a few have identified their precise locations.
In an innovative study, researchers at Stanford University in California have used Google street view images and computer vision algorithms to count the number and density of cameras in 10 major US cities and six other large cities around the world.
The study found that Seoul in South Korea had the largest number of cameras per linear kilometre (0.95), while Seattle had the lowest with 0.07 cameras per kilometre. London – which has long been touted as the CCTV capital of the world – only came in around mid-table at 0.45, surprisingly behind cities such as Paris (0.76), Boston (0.63) and San Francisco (0.52). Though, in terms of sheer numbers, London was still estimated to have around 13,000 CCTV cameras, only behind Tokyo and Seoul.
Google street view
Google street view consists of millions of 360-degree panoramas collected by cameras mounted on top of Google’s vehicles, covering more than 10 million miles across 83 countries. Using computer vision, the team at Stanford analysed 1.6 million street view images. In effect, their research method involved cameras on three levels: (i) they used a machine vision camera to analyse images of (ii) CCTV cameras, obtained from (iii) vehicle-mounted 360-degree cameras.
“When the 10 US cities were examined in more detail, researchers found that cameras tend to be concentrated in commercial, industrial and mixed city zones, rather than in residential areas. They were also more prevalent in areas with higher proportions of non-white residents, pointing to the potential impacts of surveillance technology on communities of colour.”
In order produce their estimates, the researchers built a computer vision model to detect cameras from street view images. They then applied a computer vision algorithm on a random sample of 100,000 images of each city to filter those likely to contain surveillance cameras. These images were then verified by humans. Then, by combining the geometry of the camera angle, the road network and building footprints, they were able to estimate the prevalence and locations of cameras throughout each city.
Previous attempts to estimate the number and positions of cameras have met with only limited success. The Electronic Frontier Foundation (EFF), for example, published the locations of cameras accessible by prosecutors in San Francisco, while market researchers have estimated the number of installed cameras using data on camera shipments. Neither of these approaches, however, are able to estimate the prevalence and specific locations of public and private cameras at scale.
When the 10 US cities were examined in more detail, researchers found that cameras tend to be concentrated in commercial, industrial and mixed city zones, rather than in residential areas. They were also more prevalent in areas with higher proportions of non-white residents, pointing to the potential impacts of surveillance technology on communities of colour. Interestingly, the camera estimates do not vary substantially between the two time periods considered (2011-2015 and 2016-2020), suggesting that the installation of cameras in these cities may have reached a plateau.
The team at Stanford created training and evaluation datasets for their camera detection model by taking each of the 2,660 geo-tagged cameras in San Francisco identified by EFF and pulling in the closest street view images. Manually labelling the resulting 13,240 images yielded 861 positive instances comprising 977 cameras. Many of the cameras listed by EFF appeared to be indoors or otherwise not visible from the street.
Limitations to the study
There are several limitations to the research. Firstly, it relies on surveillance cameras being captured from the street on street view images so indoor cameras – as well as outdoor cameras obscured from view – are not counted in. Secondly, due to the limits on the resolution of street view images, small cameras are difficult to detect either by algorithms or by humans, so the results are likely to underestimate the number of cameras. Thirdly, errors in the estimated recall of the computer vision model and errors in the estimated coverage of images can bias any estimates.
Finally, the method used does not provide further information about the cameras beyond the fact that they are in-situ, such as whether they are working or they are decoys. Similarly, the researchers cannot tell who owns the cameras, who has access to images and whether footage from cameras is stored – factors that are critical in assessing the downstream consequences of video surveillance.
In spite of these limitations, the researchers believe their approach and results constitute an important step toward understanding surveillance technology around the world, and the methodology could be extended and applied to a variety of other aspects of cities captured on street view images.
Sources: IFSEC global - https://www.ifsecglobal.com/video-surveillance/research-cctv-camera-prevalence-in-cities-google-street-view/?elq_mid=5980&elq_cid=1780610