In response to a freedom of information request from WIRED, TfL said it used existing CCTV images, AI algorithms and “a number of detection models” to detect behavioral patterns. “By providing station staff with insights and notifications about customer movement and behavior, it is hoped that they will be able to respond more quickly to any situation,” the response said. The company also said the trial provides insights into fare evasion that will “help inform future approaches and interventions” and that the data collected is in line with the company’s data policy.
In a statement sent after this article was published, Mandy McGregor, TfL’s director of policy and community safety, said analysis of the trial results was continuing and that the data collected from the trial included: There was no evidence of bias,” he added. McGregor said no signage was posted at the station during the trial to mention the testing of AI monitoring tools.
“We are currently evaluating the design and scope of the second phase of testing. No other decisions have been made regarding expanding the use of this technology to further stations or adding capabilities. “No,” McGregor said. “More widespread deployment of this technology beyond the pilot will require full consultation with local communities and other stakeholders, including experts in the field.”
computer vision system, It works by trying to detect objects and people in images and videos, such as those used in tests. At the London trial, an algorithm trained to detect specific actions and movements was combined with images from his CCTV camera installed 20 years ago in a subway station to detect his one in ten Images were analyzed every second. When the system detects one of 11 behaviors or events identified as problematic, it issues an alert to a station employee’s iPad or computer. According to the document, TfL staff received 19,000 potential alerts to act on, and he had a further 25,000 kept for analysis purposes.
The categories the system tried to identify are: crowd movements, unauthorized access, safeguarding, mobility aids, crime and anti-social behavior, people on the tracks, injured or unwell people, hazards such as rubbish or wet floors, abandoned items, stranded customers, and fare evasion. Each has multiple subcategories.
Daniel Rufer, senior policy analyst at digital rights group Access Now, said that whenever he sees a system that does this kind of surveillance, the first thing he looks for is whether it’s trying to detect an attack or a crime. That’s what it means. “Cameras do this by identifying body language and behavior,” he says. “What kind of dataset do I need to train something on it?”
TfL’s report on the trial said it had “wanted to include acts of aggression” but found they were “unsuccessfully detected”. It also adds that there was a lack of training data, and blacks out other reasons why acts of aggression were not included. Instead, the system would issue an alert if someone raised their arm, which the document described as a “common behavior associated with an act of aggression.”
“Training data is always insufficient because these things are probably too complex and nuanced to be properly captured in a dataset with the necessary nuances,” Leufer said, adding that TfL has He pointed out that it was positive to admit that there was no data. “I’m not sure whether machine learning systems can be used to reliably detect aggression, rather than simply reproducing existing social biases about what behavior is acceptable in public. I’m very skeptical.” According to documents received by WIRED, there were a total of 66 warnings for aggressive behavior, including test data.