Industrial DVS Design; Key Features and Applications

Event: Unknown Conference 2019 · Duration: 26 min · ▶ Watch on YouTube

Abstract

This presentation details the evolution of Samsung’s Dynamic Vision Sensor (DVS) chips, from early R&D versions to commercial products. It highlights key design features such as low latency, minimized motion artifacts, anti-flicker capabilities, and low power operation. The talk also explores practical applications of DVS technology, including sparse edge-based object recognition, human detection with privacy-preserving edge images, and pose estimation using DVS-SLAM, demonstrating its advantages over traditional CMOS Image Sensors (CIS) in challenging environments.

Speakers

  • Hyunsurk Eric Ryu — System LSI, Samsung Electronics

Talks (1)

  • 00:00 — Hyunsurk Eric Ryu: Industrial DVS Design; Key Features and Applications
    • Introduction to Samsung’s Dynamic Vision Sensor (DVS) development, its key features, and various applications including human detection and SLAM.

Key Takeaways

  • Samsung has developed multiple generations of DVS chips, progressively improving resolution, dynamic range, event processing rates, and power consumption.
  • Key design challenges in DVS include managing high event rates, minimizing AER-induced latency and motion artifacts, and suppressing unwanted ‘tail events’ and flicker.
  • DVS technology enables low-power operation and privacy-preserving human detection by processing sparse edge images, leading to smaller and faster neural networks compared to CIS-based solutions.
  • DVS-SLAM offers robust tracking even at high-speed motions and in high dynamic range scenes, outperforming CIS-based SLAM in certain challenging conditions.
  • Future work involves bandwidth minimization for higher event rates, improving pixel sensitivity in dark conditions, and developing hybrid DVS+CIS systems to leverage the strengths of both sensor types for optimal performance.

Methods / Models / Datasets Mentioned

  • DVS Gen1
  • DVS Gen2
  • DVS Gen3
  • DVS Gen4
  • G-AER (Group Address Event Representation)
  • Caffe v1
  • FRCNN (Faster Region Convolutional Neural Network)
  • FPN (Feature Pyramid Network)
  • DVS-SLAM
  • VINS (Visual-Inertial Navigation System)

Topics

Dynamic Vision Sensor (DVS) · Event-based camera · Low latency · Motion artifacts · Anti-flicker · Low power operation · Human detection · SLAM (Simultaneous Localization and Mapping) · Edge computing · Pixel design

Notes

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