Moving Vehicle Registration and Super-Resolution

Abstract

We describe a method for registering and super-resolving moving vehicles from aerial surveillance video. The challenge of vehicle super-resolution lies in the fact that vehicles may be very small and thus frame-to-frame registration does not offer enough constraints to yield registration with sub-pixel accuracy. To overcome this, we first register the large-scale image backgrounds and then, relative to the background registration, register the small-scale moving vehicle over all frames simultaneously using a vehicle motion model. To solve for the vehicle motion parameters we optimize a cost function that incorporates both vehicle appearance and background appearance consistency. Once this process accurately registers a moving vehicle, it is super-resolved. We apply both a frequency domain and a spatial domain approach. The frequency domain approach can be used when the final registered vehicle motion is well approximated by shifts in the image plane. The robust regularized spatial domain approach handles all cases of vehicle motion.