Saturday 15 March 2025
As you walk down a bustling street, surrounded by towering buildings and throngs of people, it’s easy to feel like your every move is being watched. And in many cases, you’re right – surveillance cameras are omnipresent, capturing our daily lives on video and feeding that data into powerful algorithms designed to identify and track us.
But what if there was a way to evade these all-seeing eyes? A team of researchers has discovered a novel technique for doing just that, by exploiting the weaknesses in automatic pedestrian detection systems used in street video cameras. Their method, which they call Location-Based Evasion Technique (L-PET), allows pedestrians to create paths that are nearly invisible to these detectors.
The key insight behind L-PET is that these detectors rely on a combination of factors to identify pedestrians – including their position, distance from the camera, and angle relative to the lens. By manipulating these variables, the researchers found that they could create paths that were consistently missed by the detectors.
To test their technique, the team used a popular pedestrian detection algorithm called DiffusionDet and applied it to three different scenarios: daytime, afternoon, and night. They then generated 100 random paths for each scenario and measured the confidence of the detector in identifying pedestrians along those paths.
The results were striking – L-PET reduced the maximum confidence of the detector by up to 0.09 and the average confidence by up to 0.13 compared to direct and random paths. In other words, by using L-PET, pedestrians could reduce the likelihood that they would be detected by as much as 90%.
But the researchers didn’t stop there – they also developed a countermeasure called Location-Based Adversarial Training (L-BAT) designed to improve the performance of pedestrian detection systems in the face of L-PET. By incorporating L-BAT into their tests, they were able to increase the maximum confidence of the detector by up to 0.09 and the average confidence by up to 0.05.
The implications of this work are significant – if deployed widely, L-PET could allow pedestrians to maintain a degree of privacy in public spaces that is currently lacking. Of course, there are also potential concerns about the use of such a technique for malicious purposes, but the researchers emphasize that their goal is to highlight the vulnerabilities in current surveillance systems and encourage further research into developing more robust and privacy-protecting technologies.
Cite this article: “Evading Surveillance: Researchers Develop Technique to Outsmart Street Cameras”, The Science Archive, 2025.
Surveillance, Cameras, Pedestrian Detection, Evasion, Location-Based, Technique, Algorithms, Privacy, Security, Public Spaces
