MSc dissertation proposal 2019/2020

High Speed Robot Vision with Event Based Cameras

Consumer cameras capture full images at few tens or hundreds of frames per second. While this is convenient for record and playback of videos for visualisation by humans, their use in robotics is not very efficient. First, they send massive amounts of information per second to the robot processing system, that most often contain no relevant information. Second, if a relevant event happens at a certain time, the robot can only be aware of it after the full image is acquired, transmitted and processed by the robot brain.

These facts have led to the creation of Event Based Cameras a.k.a DVS (Dynamic Vision Sensor). These cameras do not send full frames. Instead, they stream, at a very high rate, the coordinates of pixels as soon as they change their brightness level above a threshold. This is more similar to the way human vision works and allows to have latencies on the order of the microsecond to detect changes in the visual array.

Despite being a better sensor for robot vision, having a pixel stream instead of a full frame, changes the image processing paradigm, and new algorithms are being developed to compute many types of robot visual skills.

In this project, it is proposed to develop pixel stream algorithms for an important robot skill: the visual odometer. This allows robots to have a notion of their own instantaneous linear and angular positions and velocities. Being able to do this is an important step to have highly dynamic robots navigating safely in the environment (e.g. drones).

The candidate to this work must have good knowledge of image processing and vision, robotics and control. It must be comfortable in programming Matlab and c/c++ languages.