Data-Driven Feature Tracking for Event Cameras With and Without Frames.

Area of Science:

• Robotics
• Computer Vision
• Sensor Technology

Background:

• Event cameras offer high temporal resolution and motion blur resilience, ideal for low-latency feature tracking.
• Existing event camera tracking methods are often handcrafted, require extensive tuning, and lack generalization.

Purpose of the Study:

• Introduce the first data-driven feature tracker for event cameras.
• Enhance robustness and generalization for feature tracking in challenging scenarios.

Main Methods:

• Developed a data-driven feature tracker leveraging low-latency events and intensity frames.
• Implemented a novel frame attention module for information sharing across feature tracks.
• Designed dual configurations: event-only and hybrid (events + frames).

Main Results:

• Achieved robust feature tracking performance, overcoming limitations of traditional methods.
• The hybrid stereo configuration provides depth information for feature tracks.
• Demonstrated improved utility in applications like visual odometry and simultaneous localization and mapping.

Conclusions:

• The proposed data-driven approach significantly advances event camera feature tracking.
• The hybrid stereo configuration offers valuable depth perception for robotics applications.
• This work paves the way for more capable and adaptable event camera systems.