Event-Based Star Tracking for Spacecraft Attitude Estimation

Star trackers estimate spacecraft attitude from observed star positions, which is critical for communication, Earth observation, and scientific pointing. Conventional APS star trackers are limited by exposure time and frame processing cadence, making rapid motion and high-frequency pointing harder to support.

Event-based cameras are attractive because they output sparse asynchronous brightness-change events at microsecond-scale temporal resolution. For sparse star fields, most pixels remain inactive, so the sensor can offer lower latency, lower data volume, reduced motion blur, and potentially lower power consumption.

The main challenge is measurement quality under low light. Existing EBS-EKF work uses a low-light event model, a magnitude-dependent centroid correction, and a 3D EKF. The remaining gap is that the centroid offset is treated as a fixed function of brightness, even though the offset should theoretically vary with image-plane star speed.

I kept the measurement-driven philosophy of EBS-EKF instead of replacing the tracker. The implemented prototype starts from a practical event-tracking baseline: accumulate event batches, extract star centroids, initialize the field with blind astrometry, and estimate attitude batch-by-batch through a weighted Wahba solve.

The speed-aware extension is inserted between centroid extraction and attitude update. For each tracked star, the prototype estimates image-plane velocity from consecutive batches, uses centroid event count as a brightness proxy, applies a speed-aware centroid correction, and computes adaptive measurement covariance before the attitude update.

The measured centroid is treated as a biased observation rather than as a perfect star location. The correction shifts the centroid along normalized image-plane velocity using an offset function conditioned on brightness proxy and speed. In the prototype, centroid support count stands in for catalog magnitude while the full target method would use apparent magnitude and calibrated event-camera parameters.

The uncertainty model is adaptive: covariance increases when motion is fast or centroid support is weak. The intuition is that fast-moving stars and dimmer or weaker clusters should influence the attitude estimate less strongly than stable, well-supported measurements.

The project uses a staged evaluation plan. Synthetic trade-space scenarios generate controlled star-motion data for speed and brightness sweeps, helping calibrate the correction and debug failure modes before large real-data runs.

The real-data stage is designed around released EBS-EKF real-night-sky data, Earth-rotation-based validation, and public variable-speed slew observations. The current results are an initial ablation, not a final accuracy benchmark.

On one representative 3.0 s track, both the baseline and speed-aware prototype processed 372 batches and approximately 8.7M events with the same average centroid count of 38.3 centroids per batch. This keeps the comparison focused on the measurement-model change rather than a different detection count.

The speed-aware prototype changed the estimated RA, Dec, and roll trajectories, showing that the new measurement layer is active end-to-end. Runtime stayed essentially the same: mean latency decreased from 8.47 ms to 8.17 ms, max latency decreased from 124.49 ms to 118.39 ms, and budget overruns decreased from 50 to 49.

These results support the engineering viability of the module, but they are intentionally framed as preliminary. A single track demonstrates integration and runtime feasibility; it does not yet prove consistent accuracy gains across regimes.