Asynchronous Event-Inertial Odometry using a Unified Gaussian Process Regression Framework

TL;DR

This work tackles asynchronous fusion of high-temporal-resolution event camera data with IMU measurements for robust SE(3) odometry. It introduces a unified Gaussian Process regression framework with a latent-variable model that analytically integrates IMU data and couples asynchronous event-based observations within a sliding-window factor graph, implemented in two variants: CT-IMU (sparse GP prior + IMU preintegration) and GP-IMU (GP regression preintegration). Key contributions include leveraging a WNOA GP prior on SE(3), deriving GP-based preintegrated increments, and applying dynamic marginalization to maintain tractable optimization while preserving data fidelity. Experimental results on public DAVIS and MVSEC datasets show competitive accuracy against state-of-the-art synchronous methods, with GP-IMU offering potential efficiency benefits in memory- and constraint-limited scenarios. The framework lays groundwork for scalable, asynchronous event-inertial odometry, with future work targeting real-time performance through aggressive sparsification and front-end optimizations.

Abstract

Recent works have combined monocular event camera and inertial measurement unit to estimate the $SE(3)$ trajectory. However, the asynchronicity of event cameras brings a great challenge to conventional fusion algorithms. In this paper, we present an asynchronous event-inertial odometry under a unified Gaussian Process (GP) regression framework to naturally fuse asynchronous data associations and inertial measurements. A GP latent variable model is leveraged to build data-driven motion prior and acquire the analytical integration capacity. Then, asynchronous event-based feature associations and integral pseudo measurements are tightly coupled using the same GP framework. Subsequently, this fusion estimation problem is solved by underlying factor graph in a sliding-window manner. With consideration of sparsity, those historical states are marginalized orderly. A twin system is also designed for comparison, where the traditional inertial preintegration scheme is embedded in the GP-based framework to replace the GP latent variable model. Evaluations on public event-inertial datasets demonstrate the validity of both systems. Comparison experiments show competitive precision compared to the state-of-the-art synchronous scheme.

Figures (6)

• Figure 1: Illustration of the two asynchronous measurements fusing ways. Left is the Continuous-time Trajectory with standard IMU preintegration (CT-IMU) method, which queries state (green dot) on CT and fuses IMU measurements (red block) separately. Right is the Gaussian Process IMU preintegration (GP-IMU) method, relying on IMU measurements for state inference. Gray image visualizes the feature tracking.
• Figure 2: System pipeline. The whole system consists of two essential modules. The front-end process the raw events to associated feature trajectories in parallel. The back-end here uses two different GP-based ways(highlight by two blocks) fusing asynchronous measurements.
• Figure 3: Factor graph of two fusing ways. (a) is the factor graph of sparse GP prior + IMU preintegration (CT-IMU) method, (b) is the GP regression preintegration method (GP-IMU). Note the IMU state in \ref{['subsec3']} contains bias term. The ready marginal area is the marginalization order of our graph.
• Figure 4: The relative errors of the translation (top) and yaw angle (bottom) in the first 40 s using the different algorithms upon dynamic_6dof sequence.
• Figure 5: (a) Estimated trajectory of dynamic_6dof aligned with ground truth. (b) Estimated trajectory of poster_translation.