Visual-Inertial Odometry
15-DoF VIO state estimation and research on DNN-based stable visual-inertial odometry for dynamic indoor environments.
Problem
Classical visual-inertial odometry suffers from sudden covariance surges in dynamic, feature-poor indoor environments — limiting its use as a reliable odometry source for autonomous mobile robots.
Approach
My team and I developed a 15-DoF state estimation model for stable VIO with dominant bias compensation. I am extending this with a DNN-based stable VIO technique that:
- Learns IMU features and fuses them with visual features
- Uses hardware time synchronization between IMU and camera
- Alleviates sudden surges in VIO covariance estimates from classical methods
- Enables seamless switching from feature-based navigation to VIO when visual features are insufficient
Results
- Stable VIO pipeline integrated into production navigation stack
- Research direction bridging classical state estimation and learning-enhanced odometry
- Supports autonomous operation in large dynamic indoor facilities
Visual-inertial odometry experimental setup for indoor AMR navigation.
Technologies
ROS 2 · VIO · IMU · Camera · PyTorch · State estimation · Sensor synchronization