Centroiding Point-Objects With Event Cameras.

Area of Science:

• Robotics and Computer Vision
• Sensor Technology
• Signal Processing

Background:

• Event-based sensors (EBS) offer low latency and high dynamic range for tracking point-objects.
• Conventional EBS centroiding relies on Gaussian distribution assumptions and long event streams (>1s), limiting accuracy for unresolved objects and negating low-latency benefits.
• EBS circuitry introduces non-Gaussian event distributions, challenging traditional localization methods.

Purpose of the Study:

• To develop a novel method for accurate and low-latency centroiding of unresolved point-objects using EBS.
• To overcome the limitations of conventional EBS centroiding methods, specifically the Gaussian distribution assumption and long event stream requirement.
• To validate a new spatiotemporal event distribution model for improved object localization.

Main Methods:

• Derived a closed-form spatiotemporal event distribution accounting for non-Gaussian EBS effects.
• Utilized Fisher analysis to demonstrate sufficient information in short event windows (≤20 ms) for position and velocity estimation.
• Created the first EBS dataset of unresolved point-objects with subpixel ground truth using a high-speed monitor.

Main Results:

• Showed that short event windows (≤20 ms) contain adequate data for accurate localization and velocity estimation.
• A small LSTM network achieved subpixel (1 pixel) position accuracy and ±17% velocity accuracy using only 5ms of event data.
• The proposed method significantly outperforms traditional approaches in centroiding fast and dim objects.

Conclusions:

• The derived spatiotemporal event distribution enables accurate and rapid centroiding of unresolved point-objects with EBS.
• Short event windows are sufficient for precise localization and velocity estimation, leveraging the low-latency advantage of EBS.
• The developed methods and released dataset facilitate future research in real-time object tracking with event-based sensors.