Automatic Recognition of Playful Physical Activity Opportunities of the Urban Environment

We investigate deep neural networks in recognizing playful physical activity opportunities of the urban environment. Using transfer learning with a pre-trained Faster R-CNN network, we are able to train a parkour training spot detector with only a few thousand street level photographs. We utilize a simple and efficient annotation scheme that only required a few days of annotation work by parkour hobbyists, and should be easily applicable in other contexts, e.g. skateboarding. The technology is tested through parkour spot exploration and visualization experiments. To inform and motivate the technology development, we also conducted an interview study about what makes an interesting parkour spot and how parkour hobbyists find spots. Our work should be valuable for researchers and practitioners of fields like urban design and exercise video games, e.g., by providing data for a location-based game akin to
Pokémon Go, but with parkour-themed gameplay and challenges.