Comparing Representations in Tracking for Event Camera-based SLAM

Event: CVPR 2021 Workshop on Event-based Vision · Duration: 3 min · ▶ Watch on YouTube

Abstract

This work investigates the influence of different event representations on the tracking performance in event camera-based SLAM systems. It focuses on two primary representations: the Event Map (EM) used in EVO and the Time Surface Map (TS) employed in ESVO. The paper formulates the tracking problem and analyzes the characteristics of each representation, highlighting their respective advantages and disadvantages. Furthermore, an enhanced tracker is proposed that intelligently combines EM and TS, leveraging a degeneracy check to switch between them and improve robustness. Experimental results on various datasets demonstrate the performance of six tracker variations, providing insights into how these representations affect SLAM accuracy and reliability.

Speakers

  • Jianhao Jiao — The Hong Kong University of Science and Technology
  • Huaiyang Huang — The Hong Kong University of Science and Technology
  • Liang Li — The Hong Kong University of Science and Technology
  • Zhijian He — The Hong Kong University of Science and Technology
  • Yilong Zhu — The University of Hong Kong
  • Ming Liu — The Hong Kong University of Science and Technology

Talks (1)

  • 00:00:00 — Jianhao Jiao: Comparing Representations in Tracking for Event Camera-based SLAM
    • This paper presents an extensive comparison of event map and time surface map representations for event camera-based SLAM tracking and proposes an enhanced tracker combining their strengths.

Key Takeaways

  • Event Map (EM) and Time Surface Map (TS) are two distinct event representations with different characteristics, impacting tracking performance in event camera-based SLAM.
  • EM offers asynchronous output and fast generation, while TS provides synchronous output and an implicit distance field for tracking.
  • An enhanced tracker that combines EM and TS, incorporating a degeneracy check, can leverage the complementary strengths of both representations to improve robustness.
  • The choice of event representation significantly influences the reliability and accuracy of event camera-only SLAM systems, with some representations performing better under specific conditions or sequences.
  • The study provides quantitative results and visual comparisons across simulated and real-world datasets, indicating potential avenues for improving state-of-the-art event-based SLAM methods.

Methods / Models / Datasets Mentioned

  • EKF-based 6-DOF tracking
  • EKF-based depth estimation
  • Log-intensity reconstruction
  • EVO
  • Plane Sweep Methods
  • ESVO
  • Forward Projection
  • Non-linear Optimization
  • Spatio-temporal Constraint
  • Hessian Matrix

Topics

Event Camera SLAM · Event Map · Time Surface Map · Visual Odometry · Tracking Performance · Degeneracy Check · Event-based Vision · Sensor Fusion

Notes

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