MA-EVIO: A Motion-Aware Approach to Event-Based Visual-Inertial Odometry.

Area of Science:

• Robotics and Computer Vision
• Simultaneous Localization and Mapping (SLAM)

Background:

• Global Navigation Satellite System (GNSS) signals are unreliable indoors, necessitating alternative localization methods.
• Visual-Inertial Odometry (VIO) fuses camera and inertial data for pose estimation but degrades under high-speed motion and poor lighting.
• Challenges include motion blur, sensor noise, and low temporal resolution impacting VIO accuracy and robustness.

Purpose of the Study:

• To develop a robust and accurate indoor localization system overcoming VIO limitations in dynamic environments.
• To propose a motion-aware event-based VIO (MA-EVIO) system for adaptive sensor fusion and pose estimation.

Main Methods:

• Implemented a hybrid tracking strategy combining sparse feature matching and direct photometric alignment.
• Introduced motion-aware keyframe selection adapting parameters based on real-time motion classification and feature quality.
• Developed adaptive sensor fusion prioritizing event data during fast motion and RGB frames/feature tracking during slow motion.

Main Results:

• MA-EVIO demonstrated superior performance on DAVIS240c and VECtor benchmarks compared to state-of-the-art methods.
• Achieved a lower mean position error (MPE) of 0.19 on DAVIS240c and MPE/MRE of 1.19%/1.28 deg/m on VECtor.
• Outperformed EVI-SAM and PL-EVIO in challenging dynamic indoor environments.

Conclusions:

• The proposed MA-EVIO system effectively enhances indoor localization accuracy and robustness.
• Adaptive sensor fusion and motion-aware keyframe selection are crucial for handling dynamic environments.
• MA-EVIO offers a promising solution for reliable pose estimation where GNSS is unavailable.