Stereo event-based visual-inertial odometry demonstrates robust state estimation for agile platforms

Visual-inertial odometry is the problem of estimating position and orientation from a camera and an IMU without a GPS signal. It’s a well-developed field for frame cameras — there are mature open-source implementations and extensive benchmarks — but for event cameras it remains active research, because the asynchronous event stream requires rethinking how camera observations are incorporated into the state estimator.

A 2025 paper in Sensors presents a stereo event-camera VIO system that tightly couples the event stream from two synchronized event cameras with IMU measurements. The stereo baseline adds metric scale directly, removing the need for monocular scale estimation, which is one of the largest sources of drift in monocular event VIO.

The tight coupling

Tight coupling means the camera observations and IMU measurements are fused at the level of the underlying state estimate rather than sequentially — you don’t estimate pose from the camera separately and then feed it to an IMU integrator; you simultaneously optimize over a window of camera observations and IMU pre-integrations to find the state trajectory that best explains all of them. For event cameras, this means defining what a “camera observation” means in terms of events, which is non-trivial.

The authors use a continuous-time formulation where the trajectory is parameterized as a spline. Events at different times within the spline window constrain the trajectory at the corresponding temporal knots. This avoids the need to define a discrete measurement epoch — you can incorporate an event at its exact timestamp rather than binning it into the nearest frame.

The IMU integration during fast maneuvers is where the stereo event VIO shows its advantage over alternatives. An aggressive roll or pitch maneuver produces a high-angular-rate event stream that frame cameras blur and that monocular event VIO handles poorly because the scale is unobservable during pure rotation. The stereo baseline and tight IMU coupling together maintain metric scale and prevent the estimator from drifting during the maneuver.

State estimation for interceptors

An interceptor needs a state estimate of its own position and attitude to close the guidance loop. GPS provides this in the pre-terminal phase. In the terminal phase — below a few hundred meters from the target — GPS may be jammed, unavailable due to update rate limitations, or simply too noisy for the required guidance bandwidth.

INS integration provides short-term state estimation without GPS, but IMU drift accumulates rapidly, especially for low-cost sensors on small platforms. Aiding the IMU with a visual odometry system that doesn’t require GPS extends the accurate navigation window. The critical property needed for terminal guidance is that the visual odometry doesn’t break during the high-angular-rate maneuvers that occur as the interceptor closes on a maneuvering target.

This is exactly the regime where frame-camera VIO fails and event-camera VIO succeeds. The stereo event VIO paper demonstrates this property explicitly — the benchmark includes sequences with angular rates exceeding 500 degrees per second, where frame-camera methods produce large errors and the stereo event system maintains metric accuracy.

Implementation path

The algorithm in the paper is computationally demanding — the continuous-time spline optimization is an iterative nonlinear least-squares problem. The authors run it on a laptop-class processor, which is not a viable flight computer. Reducing the computational cost requires either reducing the spline resolution (accepting more approximation error) or using a sliding-window linearization that reduces the optimization problem size at the cost of some accuracy.

For airborne deployment, the more practical near-term path may be a looser coupling: use the event-camera stereo for feature tracking during normal flight and fall back to pure IMU integration during high-angular-rate maneuvers, with the event camera providing a correction when conditions allow. This is less elegant than full tight coupling but is implementable on current embedded processors and may be sufficient for the guidance accuracy requirements.

The stereo event VIO work is important because it demonstrates that the state estimation problem for agile platforms is solvable with current sensor technology. The gap between algorithm performance and embedded deployment remains, but it’s an engineering gap rather than a fundamental limitation.