Mix-VIO: A Visual Inertial Odometry Based on a Hybrid Tracking Strategy

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Conventional visual odometry struggles with dynamic lighting and image blur.
• Robust visual-inertial odometry (VIO) is crucial for autonomous systems.

Purpose of the Study:

• To propose Mix-VIO, a novel hybrid visual-inertial odometry system.
• To enhance front-end tracking performance under challenging environmental conditions.

Main Methods:

• Implemented a hybrid tracking approach combining handcrafted and Deep Neural Network (DNN)-based techniques.
• Utilized DNN for feature point detection and integrated optical flow and DNN-based sparse feature matching.
• Employed sliding window and bundle adjustment (BA) for back-end optimization and pose estimation.

Main Results:

• Mix-VIO demonstrated robust tracking performance under dynamic lighting.
• The system maintained optimal results even with significant image blur.
• Experimental validation confirmed the effectiveness of the hybrid feature extraction and matching methods.

Conclusions:

• Mix-VIO offers a significant improvement over conventional VIO systems.
• The hybrid approach effectively addresses limitations in dynamic and blurred environments.
• This work advances the reliability of VIO for real-world applications.