Towards Asynchronous SLAM with Event Cameras

Event: CVPRW 2021 · Duration: 15 min · ▶ Watch on YouTube

Abstract

Event cameras offer significant advantages over traditional cameras for robotics applications, including fast perception, high dynamic range (HDR) capabilities, and low power consumption, by producing asynchronous and sparse event streams. This presentation outlines the development of an asynchronous event-driven SLAM pipeline, focusing on novel methods for feature detection and tracking that are specifically designed to leverage the unique spatio-temporal nature of event data. The speaker discusses how these event-driven algorithms can overcome limitations of traditional frame-based approaches, enabling robust and real-time performance in challenging scenarios.

Speakers

  • Ignacio Alzugaray — Vision for Robotics Lab, ETH Zurich

Talks (1)

  • 00:00:00 — Ignacio Alzugaray: Towards Asynchronous SLAM with Event Cameras
    • This talk introduces the challenges and opportunities of using event cameras for asynchronous SLAM, detailing event-driven feature detection and multi-hypothesis tracking methods that exploit the unique properties of event data.

Key Takeaways

  • Event cameras provide fast, low-power, and HDR perception, generating asynchronous and sparse data streams that require specialized event-driven algorithms.
  • Traditional SLAM pipelines, designed for frame-based cameras, need significant adaptation to effectively utilize event camera data, moving from discrete image processing to continuous event-by-event updates.
  • Asynchronous corner detection and multi-hypothesis tracking methods have been developed to robustly extract and track features directly from event streams, enabling high-speed tracking even under challenging motion and illumination conditions.
  • The HASTE framework introduces an incremental alignment score update, significantly improving computational efficiency and achieving real-time performance for event-driven feature tracking.
  • Event-driven perception is a natural fit for event cameras, explicitly exploiting data sparsity and asynchronicity, and reducing assumptions about motion speed or texture density, though requiring careful algorithm design for efficiency, robustness, and scalability.

Methods / Models / Datasets Mentioned

  • Multi-Agent Visual SLAM
  • Vision-based Navigation & Manipulation
  • Viewpoint-tolerant Place Recognition
  • Surface of Active Events (SAE)
  • Asynchronous Corner Detection and Tracking for Event Cameras in Real-Time [Alzugary & Chili, RAL'18]
  • ACE: An Efficient Asynchronous Corner Tracker for Event Cameras [Alzugary & Chili, 3DV'18]
  • Asynchronous Multi-Hypothesis Tracking of Features with Event Cameras [Alzugary & Chili, 3DV'19]
  • HASTE: Multi-Hypothesis Asynchronous Speeded-up Tracking of Events [Alzugary & Chili, BMVC'20]
  • Frame-like, event-based tracking [Zhu et al., ICRA'17]

Topics

Event cameras · Asynchronous SLAM · Feature detection · Corner tracking · Multi-hypothesis tracking · Event-driven perception · Robotics applications · Real-time performance · Sparsity · High Dynamic Range (HDR)

Notes

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