Data Collection: The Simplicity of Patterns vs. the Complexity of Situations

11.06.2024|Christian Kreutz

Driverless trains have been operational for many years, at least on short routes. A rules-based algorithm can efficiently control a train due to its predictable movement patterns. However, autonomous car driving presents a entirely different challenge, as unforeseen situations not accounted for in the model can arise at any moment. In San Francisco, autonomous cars even blocked access when confronted with an unexpected emergency situation they couldn't navigate.

Wired magazine features a captivating article about a metro security surveillance project in London. Over the course of one year, the public transport authority tested in one station 11 artificial intelligence models designed to detect behavioral violations captured by surveillance cameras, such as ticket control breaches, vandalism, or personal attacks.

Detecting someone jumping over the ticket control counter is relatively simple, but identifying "wrong" behavior is a far more complex task. Imagine a person momentarily placing a piece of paper on the wall to write something down—this could trigger a false graffiti alarm. Similarly, teenagers dancing and playfully pushing each other while listening to music could set off a false gang violence alert.

During the one-year test, more than 44,000 alerts were generated, with 19,000 being delivered to station staff in real-time. This translates to between 50 and 120 alerts per day for a single metro station. It would be interesting to investigate the workload implications of such a high volume of alerts, especially considering that London has over 250 metro stations.

Consequently, numerous false alarms occurred. To enhance the system's pattern recognition capabilities, they discovered that many potential scenarios were not incorporated into the model. Ironically, at one point, security personnel posed with weapons in front of the cameras to train the models. It would be fascinating to learn about the types of scenes they enacted to ensure realistic representation.

The question remains whether this approach will genuinely enhance metro station security and enable police officers to respond promptly. In a world with an increasing number of cameras, their primary use seems to be documenting crime rather than preventing it.