Project Overview
At Aegis Rider, I worked on augmented reality helmets designed for dynamic moving systems such as motorcycles and cars. The core problem was maintaining reliable alignment between the digital overlay and the physical world while both the rider and vehicle were moving aggressively.
My contribution centered on a visual-inertial tracking pipeline built around an Extended Kalman Filter. The system estimated both head pose and vehicle pose in real time, which made this project a tight combination of sensor fusion, estimation quality, calibration discipline, and low-latency systems engineering.
Technical Architecture
Visual-Inertial Odometry
Fused camera and motion signals into a consistent estimate of motion under aggressive dynamics
Extended Kalman Filter
Used an EKF-based estimation pipeline to stabilize pose tracking for both rider and vehicle motion
AR Alignment
Maintained spatial consistency between the physical scene and the rendered experience
Real-Time Systems
Optimized for responsive behavior in a product environment where latency directly affected usability
What Made the Project Hard
This was not a standard AR demo. The system had to stay stable when both the user and platform were moving unpredictably:
- Aggressive motion: Tracking had to remain usable even when vibration, acceleration, and rapid orientation changes were present
- Multi-frame consistency: Small drift or timing errors quickly become obvious in AR, so state estimation quality mattered continuously
- Latency pressure: The perception stack had to deliver useful estimates quickly enough for an interactive wearable product
- Calibration discipline: Robust deployment depended on getting sensor models, synchronization, and integration details right
Implementation Focus
Pose estimation under motion: The tracking stack had to make sense of rapid motion without introducing instability into the AR experience. That meant paying close attention to estimator behavior, failure modes, and temporal consistency.
Head and vehicle frame reasoning: The problem involved more than just estimating one moving body. Reliable output depended on reasoning across both head motion and vehicle motion in a coordinated real-time pipeline.
Product readiness: This was estimation work in a product context, which changes the bar. The solution had to be not only technically correct, but predictable, maintainable, and usable under real operating conditions.
Impact
This project deepened my practical experience in visual-inertial estimation, real-time tracking, and the gap between a technically interesting pipeline and a product-ready perception system.
It also reinforced a theme that carries through much of my work: the best perception systems are the ones that stay trustworthy when the operating conditions stop being friendly.