Event-based Algorithms for Robust and High-speed Robotics

Event: CVPR 2017 Workshop on Event-based Vision · Duration: 32 min · ▶ Watch on YouTube

Abstract

The talk highlights the limitations of traditional frame-based cameras in high-speed and high-dynamic-range scenarios for robotics, advocating for event-based cameras as a revolutionary alternative. It presents several research areas including visual-inertial state estimation, vision-based navigation for flying robots, deep learning for end-to-end navigation, and low-latency vision for aggressive flight. The core of the presentation focuses on event-based algorithms for 6-DOF pose tracking and parallel tracking and mapping (PTAM), demonstrating their robustness to motion blur and high dynamic range. The speaker also discusses continuous-time trajectory estimation and real-time visual-inertial odometry, emphasizing the potential for a two-level sensing architecture combining fast event-based sensors with slower cognitive sensors.

Speakers

  • Davide Scaramuzza — Robotics & Perception Group, University of Zurich

Talks (1)

  • 00:00 — Davide Scaramuzza: Event-based Algorithms for Robust and High-speed Robotics
    • This presentation introduces event-based cameras as a solution for low-latency perception and control in high-speed robotics, showcasing various algorithms for pose tracking, mapping, and visual-inertial odometry in challenging scenarios.

Key Takeaways

  • Event-based cameras offer significant advantages over traditional frame-based cameras for high-speed and high-dynamic-range robotic applications due to their low latency and robustness to motion blur.
  • Algorithms like EVO demonstrate real-time 6-DOF pose tracking and mapping capabilities using event data, even on resource-constrained platforms like smartphone CPUs.
  • Continuous-time trajectory estimation using B-splines and non-linear optimization allows for robust fusion of event camera and IMU data, providing accurate pose estimation.
  • A future two-level sensing architecture is proposed, combining fast event-based sensors for agile behavior with slower traditional sensors for cognitive tasks like recognition and mapping.
  • A publicly available event camera dataset and simulator are provided to foster further research and development in event-based vision and SLAM.

Methods / Models / Datasets Mentioned

  • SVO
  • DTAM
  • REMODE
  • EVO
  • B-splines
  • PTAM
  • Odroid XU4

Topics

Event-based cameras · High-speed robotics · Low-latency perception · Pose tracking · Visual-inertial odometry · SLAM · Motion blur · High dynamic range · Drone navigation · Sensor fusion

Notes

Open for commentary — connections to other work, critiques, follow-up reading.