Hybrid-EVIO: Event-Based Visual-Inertial Odometry With Hybrid Visual Front-End

Frame-based visual-inertial odometry (VIO) can be significantly enhanced by integrating event cameras, particularly in high-speed motion and high-dynamic-range (HDR) scenarios. However, existing VIO frameworks often process event and frame-based data simultaneously, introducing unnecessary computational overhead. Additionally, tracking handcrafted features on event streams typically requires extensive parameter tuning and lacks robustness against noise. To address these limitations, we propose Hybrid-EVIO, a method that effectively fuses event data, standard frames, and inertial measurement unit (IMU) measurements. Hybrid-EVIO consists of a hybrid visual front-end and a sliding-window-based back-end. The front-end combines traditional and learning-based techniques in a scene-adaptive manner: features are tracked using either conventional methods on frames or a learned sparse optical flow network on asynchronous events, depending on the imaging quality. IMU measurements are further utilized to construct epipolar constraints, prefiltering extreme outliers before Random Sample Consensus and thereby improving pose estimation accuracy. Finally, a tightly coupled, graph-based optimization framework integrates three sensor modalities for high-precision state estimation. We evaluate the proposed method on multiple representative and challenging public datasets. Our approach outperforms state-of-the-art methods, reducing trajectory errors by up to 34% in the best case. Our trajectories and evaluation code are publicly available at https://github.com/sssxxxkkkk/hybrid EVIO.