Motion Estimation by Focus Optimisation: Optic Flow and Motion Segmentation with Event Cameras

Abstract

Event cameras are a new camera technology inspired by biological vision. Instead of producing frames like a normal camera, each pixel is able to send neural spikes (called events) independently and asynchronously, just like the visual receptors in your eye. This has huge advantages for robot vision, since event cameras are fast, low power, don t have motion blur, and have high dynamic range. Processing such data is a challenge that this thesis helps overcome, by using cutting edge artificial neural networks, machine learning, and optimisation techniques to estimate motion and reconstruct video with near-infinite framerate from the events.