Lightweight event camera inertial fusion pose estimation with Cramer Rao lower bound analysis

To address the limitations of traditional frame-based PnP algorithms—such as motion blur, limited dynamic range, and high computational complexity under high-speed motion and extreme illumination changes—this paper proposes the Event-PNP algorithm framework. This framework tightly couples asynchronous event streams with IMU data at the measurement model level, enabling high-precision real-time six-degree-of-freedom pose estimation. The algorithm establishes an event geometry model based on luminance constancy constraints, combining event triggering with pixel motion to avoid the temporal resolution loss inherent in traditional event frame accumulation. It employs IMU pre-integration techniques to jointly minimize event geometry errors and IMU residuals within a unified optimization framework. Through lightweight strategies such as incremental optimization and sparsification, the algorithm’s computational complexity is reduced to linear-time. Theoretically, an event-inertial fusion Fisher information matrix framework is established, deriving a closed-form lower bound expression for the Cramer-Rao inequality to provide a theoretical benchmark for performance evaluation. Experiments demonstrate that Event-PNP achieves rotational accuracy of 0.08° and translational accuracy of 1.2 cm, representing 35% and 42% improvements over existing state-of-the-art methods. On embedded platforms, it operates at update rates exceeding 200 Hz with a statistical efficiency of 0.89. The algorithm maintains stable tracking under extreme conditions such as high-speed motion (520°/s) and stroboscopic lighting (100 Hz), demonstrating outstanding performance in applications like AR registration and drone landing. This work lays the theoretical and technical foundation for the large-scale application of event cameras in high-dynamic visual localization.