Aerial Car Detection and Urban Understanding

In this work we investigate car detection from aerial imagery and explore how it can be applied to urban understanding. To perform car detection we use the rotationally-invariant Fourier HOG detector. By adding incremental changes we are able to improve its detection probability by 10% for a range of false alarm rates. Further improvements can be made if we filter out cars that are not near known streets or inside car parks. We use the detected cars for automatic urban understanding: street estimation, car park detection and monitoring. In our experiments we were able to detect about half of all car parks in two major cities. Our method for car park monitoring allows us to find simple trends in car park usage, as well as changes in car park structure. We expect this information to be highly useful for future city planning.