Researchers Discover How to Hide People from Surveillance Cameras with 2D Images
Researchers from KU Leuven in Belgium have published a scientific report on deceiving surveillance systems. It turns out that a simple 2D image, printed on a T-shirt or bag, can make a person invisible to surveillance cameras that rely on machine learning to detect people in video streams.
How the Method Works
To achieve the desired effect, a 40×40 centimeter image (referred to as a “patch” in the report) needs to be placed in the center of the camera’s detection box and remain in its field of view at all times. While this method does not hide a person’s face, the human detection algorithm will not recognize a person in the frame, which means facial recognition will not be triggered either.
Experiment Results
During their experiments, the researchers tried various images to fool the surveillance systems, including abstract “noise” and blurred pictures. However, they found that photos of random objects, processed in different ways, worked best. For example, the patches shown in the illustration below were created from random images that were rotated by 20 degrees, randomly scaled, had noise added, and had their brightness and contrast randomly modified.
When these images are printed on clothing, bags, and similar items, the algorithms stop detecting the person behind them. You can see the effectiveness of these images in a proof-of-concept video published by the researchers. They tested their method on the open-source neural network Darknet, which uses the YOLOv2 (You Only Look Once) real-time object detection system.
Applications and Open Source Code
This method can be used not only to hide people from cameras, but also any other object. For example, a surveillance system will not “see” a car or a bag with a patch applied to it.
In addition to publishing their report and video, the researchers have also uploaded the source code they used to create the patches on GitHub, so anyone interested can replicate and continue their experiments.