Reconstruction, Motion Estimation and SLAM from Events

Event: ICRA 2017 · Duration: 34 min · ▶ Watch on YouTube

Abstract

This presentation explores the evolution of Visual SLAM, from its early focus on sparse localization and mapping to current advancements in dense and semantic understanding. The speaker emphasizes the increasing maturity of SLAM systems, as evidenced by their integration into commercial products, while also highlighting the ongoing challenges in achieving real-world robustness, low power consumption, and compactness for embedded AI and Intelligence Augmentation (IA) applications. A significant portion of the talk is dedicated to introducing event-based vision as a novel paradigm that addresses many of these challenges, offering high-speed, low-latency, and high-dynamic-range data for real-time reconstruction, motion estimation, and SLAM. The presentation concludes by outlining future directions, including the co-design of sensors, processors, and algorithms, and the potential of neuromorphic computing to unlock new levels of performance and efficiency in spatial perception.

Speakers

  • Andrew Davison — Robot Vision Group and Dyson Robotics Laboratory, Department of Computing, Imperial College London

Talks (1)

  • 00:00:00 — Andrew Davison: Reconstruction, Motion Estimation and SLAM from Events
    • A presentation on the evolution of Visual SLAM, from sparse mapping to dense and semantic understanding, highlighting the need for efficient, robust, and useful systems for embedded AI/IA applications, and introducing event-based vision as a promising direction for future SLAM systems.

Key Takeaways

  • SLAM is evolving into a general real-time spatial perception capability, essential for embedded AI and Intelligence Augmentation (IA) applications, requiring robust and useful systems.
  • Modern SLAM research is moving beyond basic localization to dense and semantic mapping, aiming to provide a deeper understanding of the environment and its objects.
  • While increasing computational power has enabled complex computer vision algorithms in real-time, embedded applications demand low power, compactness, and real-world robustness, necessitating a shift in design philosophy.
  • Event-based vision offers a promising path forward for SLAM, providing data that is purely event-based, minimizes latency, and is well-suited for future integrated sensor/processor architectures.
  • The future of real-time vision involves the co-design of algorithms, processors, and sensors, potentially leveraging novel architectures like graph processors and neuromorphic principles to achieve the required performance and efficiency for widespread adoption.

Methods / Models / Datasets Mentioned

  • MonoSLAM
  • Dyson 360 Eye
  • Google Project Tango
  • Microsoft HoloLens
  • DTAM (Dense Tracking and Mapping)
  • Height Map Fusion with Dynamic Level of Detail
  • SemanticFusion
  • SLAM++
  • SLAMBench (PAMELA Project)
  • Graphcore's IPU (Intelligent Processing Unit)
  • IBM TrueNorth
  • Brainchip
  • DVS128 (Dynamic Vision Sensor)
  • Particle Filter
  • Pixel-wise EKF (Extended Kalman Filter)
  • Optical Flow Constraint

Topics

Visual SLAM · Event-based Vision · Real-time Spatial Perception · Dense Mapping · Semantic Understanding · Embedded AI/IA · Neuromorphic Computing · Sensor-Processor Co-design · Motion Estimation · 3D Reconstruction

Notes

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