An international airline.
For airlines, delays are anathema: they are expensive, damaging to the brand and inconvenient for customers. One cause of delays is the time it takes for passengers to disembark. If the disembarkation process could be made as efficient as possible, delays would be reduced. The airline asked Faculty to research two questions: typically, how long does disembarkation take, and could we measure the flow of people over the course of disembarkation?
We used the footage supplied by the airline of disembarkations as data to train an image recognition model to review it for us. First, the model needed to identify individuals and their luggage in the footage. To do this we used a conventional object detector termed Mask R-CNN, which enabled us to track individuals even in a crowded scene. Second, the model tracked individuals across the CCTV frames in order to measure the time it took them to complete the whole process of disembarkation. We combined two techniques to achieve this: estimating the location of the object (the individual) in the following frame and using this to verify the objects as one and the same; and using a ‘visual similarity metric’ to ensure that the same individual is being tracked throughout.
Once tracked objects are combined, a clear pattern emerges that matches what we might expect. This new information on how passengers and crew disembark offers the airline a new opportunity to tackle costly delays.
Passenger flow over time. In the red region the majority of passengers disembark, in grey the passengers in need of assistance disembark, and finally cabin crew in blue.