New York’s MTA Strapped Google Pixels To Moving Subway Cars, For Science

In partnership with Google, the Metro Transit Authority took to tying several Google Pixel phones to subway cars in New York in order to track data. The experiment was part of the MTA’s need to automate and expand track-safety inspections and repair, and tying a few Pixels to a train car seemed the best place to start.

If you’re into facts, the New York City subway carries somewhere over 3 million people per day, according to its own website. With that sort of volume, human-powered efforts to find, track defects, and maintain them can only be so efficient. It works, but there’s room for improvement.

The MTA is looking to supplement that effort with some sort of automation, and AI partnered with existing technology isn’t so hard to get your hands on. According to WIRED, Google and the New York MTA partnered up to strap several devices to subway cars in order to listen for track defects while recording other movement data. Those devices weren’t some specialized hardware for the professional sector; they were Pixels.

The Google Public Sector worked with the New York City subway to provide several Pixels in an effort to experiment with existing, off-the-shelf hardware under operation TrackInspect. If TrackInspect was a success, it would showcase that everyday phones have the capability to provide enough data to supplement the work done by individuals in repairing and maintaining the rail system. The Pixels would need to collect audio, movement, and geographic data underground to be fed to AI training models that could efficiently package the data for repair teams. All of the sounds commuters take for granted — the screeches and heavy crashes or bumps — could be translated into a specific track that needs attention.

While human inspection is still required, the goal is to automate most of the flagging system. Through all of the recordings of the New York subway that were made with the stowaway Pixel phones, 92% of recorded defects were corroborated by human inspectors. This project still used an inspector to listen to all of the collected audio and analyze vibration recordings, with an 80% success rate. The TrackInspect project collected 335 million sensor readings and 1,200 hours of audio. Those collections were used to train around 200 individual AI models for this exact work.

The hope is that the MTA will be able to implement this technology further, possibly with specialized hardware instead of a device that was made to do something entirely different. The project proves that the tech available now can be implemented at little cost with the proper AI models involved.

Add 9to5Google to your Google News feed.