Title:

Event-based optical flow estimation, event-based depth map estimation and LiDAR. Application to outdoor driving scenes.

Abstract:

Event-based cameras are nowadays very popular sensors in industry, in various fields such as machine vision, predictive maintenance, augmented reality or driver assistance. They must facilitate the analysis of complex dynamic scenes made up of rapidly moving objects, under lighting conditions that can vary greatly.
In this seminar, I will present a real-time optical flow calculation method adapted to this type of camera. A new dense representation of the sparse event stream, in the form of an "inverse exponential distance surface", allows us to reuse state-of-the-art image frame-based optical flow computation methods. I will also present a second architecture that allows us to learn to estimate two dense depth maps from LiDAR and event data. These two maps allow us to associate each event with a depth difference.

Bio:

Franck Davoine is a CNRS researcher at the Heudiasyc laboratory, UMR CNRS / Université de technologie de Compiègne.