In a rather bizarre incident, a Dutch man has been charged 380 Euros (about $400) for purportedly using his cell phone while driving, although he insists that he was just scratching his head. Tim Hansen received the penalty when a Monocam system of an AI camera took a snap of an image appearing to depict him conversing on his mobile phone. Yet upon further scrutiny, it turned out that Hansen did not, in fact, hold anything in his hand – he was merely itching the side of his head.
Hansen, an IT specialist who designs algorithms for image manipulation and analysis purposes, drew upon his knowledge to illustrate how Monocam or similar systems could err. He clarified that these systems learn from a vast collection of images split into training, validation, and test sets. The training set helps the algorithm understand objects and their characteristics in the images, while the validation set tunes the algorithm’s parameters. Finally, the test set gauges the system’s performance.
Hansen claims that Monocam might have erred in that its training dataset most likely includes a large number of images of people putting phones to their ears but few or none of people holding empty hands to their ears. This suggests that rather than relying on the existence of a phone, the algorithm might have trained to recognize a phone based on hand position.
The instance of Hansen demonstrates the shortcomings of AI-driven camera systems in precisely detecting infractions. Errors can still happen even with efforts to reduce false positives, such as human image inspection. Hansen is awaiting an official ruling on his fine dispute, which may take up to 26 weeks.
The dependability of these kinds of camera systems has come under scrutiny in the Netherlands and other surrounding nations as a result of this incident. While some organizations are pushing for the installation of cameras that can identify when someone is using a phone while driving, Hansen’s experience serves as a warning that human oversight may still be required because these technologies are not perfect.