Event-Based Algorithms For Geometric Computer Vision

Loading...
Thumbnail Image
Degree type
Doctor of Philosophy (PhD)
Graduate group
Computer and Information Science
Discipline
Subject
event-based cameras
event cameras
self-supervised learning
Artificial Intelligence and Robotics
Computer Sciences
Robotics
Funder
Grant number
License
Copyright date
2020-02-07T20:19:00-08:00
Distributor
Related resources
Author
Zhu, Alex Zihao
Contributor
Abstract

Event cameras are novel bio-inspired sensors which mimic the function of the human retina. Rather than directly capturing intensities to form synchronous images as in traditional cameras, event cameras asynchronously detect changes in log image intensity. When such a change is detected at a given pixel, the change is immediately sent to the host computer, where each event consists of the x,y pixel position of the change, a timestamp, accurate to tens of microseconds, and a polarity, indicating whether the pixel got brighter or darker. These cameras provide a number of useful benefits over traditional cameras, including the ability to track extremely fast motions, high dynamic range, and low power consumption. However, with a new sensing modality comes the need to develop novel algorithms. As these cameras do not capture photometric intensities, novel loss functions must be developed to replace the photoconsistency assumption which serves as the backbone of many classical computer vision algorithms. In addition, the relative novelty of these sensors means that there does not exist the wealth of data available for traditional images with which we can train learning based methods such as deep neural networks. In this work, we address both of these issues with two foundational principles. First, we show that the motion blur induced when the events are projected into the 2D image plane can be used as a suitable substitute for the classical photometric loss function. Second, we develop self-supervised learning methods which allow us to train convolutional neural networks to estimate motion without any labeled training data. We apply these principles to solve classical perception problems such as feature tracking, visual inertial odometry, optical flow and stereo depth estimation, as well as recognition tasks such as object detection and human pose estimation. We show that these solutions are able to utilize the benefits of event cameras, allowing us to operate in fast moving scenes with challenging lighting which would be incredibly difficult for traditional cameras.

Advisor
Kostas . Daniilidis
Date of degree
2019-01-01
Date Range for Data Collection (Start Date)
Date Range for Data Collection (End Date)
Digital Object Identifier
Series name and number
Volume number
Issue number
Publisher
Publisher DOI
Journal Issue
Comments
Recommended citation