Source-linked AI summary
Spacecraft Pose Tracking via Kinematic Prediction and Selective Event Correction
Yonglong Zhang, Guanning Yao, Shuo Zhang, Yifan Yang, Yang Liu†, Zongwu Xie
TL;DR
Initialized spacecraft tracking can use recent pose history to predict motion while event support changes, but correction must select which predicted components to update. OA-EI applies scaled, activity-supported event corrections to rotation and image-plane translation while retaining predicted depth. On the evaluated FRESH cohort, it reduces 7.7-second rotation and translation errors relative to kinematic prediction.
Problem
Event support varies in amount and spatial coverage, while initialized spacecraft tracking needs continuous orientation and translation estimates and can use recent pose history for short-horizon prediction.
Source: §I · p. 1Method
OA-EI propagates a complete pose with constant-body-twist motion and uses event observations to correct rotation and normalized image-plane translation while retaining predicted axial depth.
Source: §I · p. 1Results
On the evaluated FRESH cohort, 7.7-s rotation and translation errors are reduced by 62.2% and 60.9% relative to kinematic prediction, and failure rates decrease from 20.0% to 1.67%.
Source: §V · p. 6Takeaways & Limitations
The evaluated state-update comparison favors rotation–bearing correction for overall pose accuracy, and combined Scale Gate and activity support outperform full injection at full, 10%, and 1% event retention.
Source: §V · p. 6The selected correction subspace retains axial depth from prediction and does not use all depth information carried by extended-object scale and temporal geometry.
Source: B. Observation-Aligned Correction · p. 3
Abstract
from arXiv · showhide
In initialized spacecraft tracking, a short-horizon motion prediction remains available as event support changes. Observation-Aligned Event Innovation (OA-EI) uses events to correct selected components of this prediction. A constant-body-twist model propagates the complete pose; the event encoder proposes rotational and normalized image-plane translation corrections, retaining propagated axial depth. Candidate magnitudes come from a bounded Scale Gate and a deterministic activity factor, with empty windows returning the prediction unchanged. Evaluation uses four FRESH physical-event trajectories, twenty rollout starts, and three training seeds. Each rollout starts from two reference poses and then uses self-estimates. At 7.7 seconds, rotation and translation errors fall from 41.63{\deg}/57.28 mm for prediction alone to 15.73{\deg}/22.41 mm for OA-EI, reductions of 62.2% and 60.9%. Failure rates fall from 20.0% to 1.67%. Matched ablations favor the rotation-bearing interface for overall pose accuracy; combining scaling with activity support gives the lowest errors at every tested event-retention level. On Jetson Orin Nano, the measured in-memory update takes 44.56 milliseconds at the median and 58.77 milliseconds at P95.
I. INTRODUCTION
Spacecraft tracking must maintain orientation and translation estimates as event quantity and spatial coverage vary, while initialized trackers can use recent pose history to predict the next state. OA-EI corrects selected parts of that prediction, and the evaluation examines recursive tracking and matched state-update and scaling ablations.
I. INTRODUCTION
- Event support varies in amount and spatial coverage, while continuous orientation and translation estimates are needed for spacecraft navigation, inspection, and servicing.Sparse observations can coincide with motion that remains predictable over a short interval.Source: §I · p. 1
- Pose history enables a short-horizon prediction before current events are used, but correction choices and magnitudes affect the history used by later predictions.The evaluation compares prediction alone, full candidate injection, and OA-EI under a common recursive protocol, with matched ablations of state updates and injection scaling.Source: §I · p. 1
- OA-EI corrects rotation and normalized image-plane translation while retaining axial depth from a complete constant-body-twist pose prediction.The event branch also uses learned correction scales and event-activity support; an empty event window returns the prediction exactly.Source: §I · pp. 1–2
A. Spacecraft Pose Sensing
- Image-based spacecraft pose methods include direct regression and keypoint-based PnP, while event-based work includes keypoint estimation with temporal smoothing or optimization.SPEED and SPEED+ benchmark image-based estimation,; SPADES and FRESH extend evaluation to synthetic and physical event recordings [4],.Source: A. Spacecraft Pose Sensing · p. 2
B. Event-Based Motion Processing
Prior event-based motion methods recover motion or pose from event streams, while prediction-and-observation estimators combine a motion prediction with learned or measured observations. OA-EI instead places learned corrections in selected components of a kinematic pose prior and feeds the corrected pose into subsequent predictions.
B. Event-Based Motion Processing
- Event-based motion processing includes event alignment, optical flow, geometric and learned odometry, and six-degree-of-freedom tracking with events and depth.OA-EI uses polarity-separated voxels and normalized image-plane translation for target-relative correction, with axial depth supplied by the predictor.Source: B. Event-Based Motion Processing · p. 2
C. Prediction and Observation Updates
- Spacecraft prediction-and-observation methods include adaptive Kalman filters, recurrent learned filtering, and latent-space prediction and fusion; OA-EI applies learned correction in a selected pose subspace.Its rotational and bearing-space candidates have separate scale coefficients, with event activity providing an additional support factor.Source: C. Prediction and Observation Updates · p. 2
III. METHOD
- OA-EI forms a complete pose prediction from recent estimates, applies rotational and image-plane corrections using a causal event window, and feeds the corrected pose into the next history.The injected correction depends on learned scaling and current event activity.Source: §III · p. 2, C. Prediction and Observation Updates · p. 2
A. Kinematics-Guided Pose Prior
- The kinematic prior propagates the historical motion trend, while unmodeled acceleration and accumulated history errors motivate correction from current events.The pose is represented by rotation and translation, including image-plane coordinates and axial depth.Source: A. Kinematics-Guided Pose Prior · p. 2
B. Observation-Aligned Correction
OA-EI corrects selected components of a propagated pose using event-derived rotational and image-plane translation candidates. It retains the prior axial depth, so translation correction changes bearing rather than depth.
B. Observation-Aligned Correction
- A causal event window supplies pixel locations, timestamps, and polarities for the correction update.The window contains events from tk+1 − h < τi ≤ tk+1.Source: B. Observation-Aligned Correction · p. 2
- Polarity-separated voxelization and a shared encoder produce a feature used to predict rotational and image-plane translation candidates.The encoder feature is 512-dimensional.Source: B. Observation-Aligned Correction · p. 3
- The prior pose supplies the residual reference and motion scalars used by the event candidates and Scale Gate.The event branch predicts rotational and image-plane translation candidates relative to the prior.Source: B. Observation-Aligned Correction · p. 3
- The prior translation is projected to normalized image-plane coordinates, where the event update changes bearing while retaining axial depth.The corrected translation is reconstructed from the updated image-plane coordinates and the prior depth.Source: B. Observation-Aligned Correction · p. 3
- OA-EI retains propagated axial depth while correcting translation through image-plane bearing, a subspace that does not capture depth information from extended-object scale or temporal geometry.Both prior and corrected translations lie on the same axial-depth plane.Source: B. Observation-Aligned Correction · p. 3, Figure 2 · p. 3
C. Learned and Activity-Supported Injection
Learned scale coefficients and a deterministic event-activity factor set how much of each candidate enters the pose update. Empty event windows leave the prediction unchanged.
C. Learned and Activity-Supported Injection
- The Scale Gate uses event features and nine causal scalars, including event count, spatial activity, prior speeds, and timing.The scalar inputs are normalized using training-set statistics.Source: C. Learned and Activity-Supported… · p. 3
- The Scale Gate is trained after the candidate network is frozen, with targets minimizing rotation geodesic and translation errors along proposed directions.Near-zero candidate directions are excluded from the masked Smooth-L1 scale loss, and nonnegative coefficients preserve candidate directions.Source: C. Learned and Activity-Supported… · p. 3
- Rotation scale targets use grid search with local refinement, while translation scale targets have an analytic form.The translation target uses the prior-depth-scaled image-plane candidate and the reference image-plane residual.Source: C. Learned and Activity-Supported… · p. 3
- The deterministic Evidence factor uses the smaller of count- and active-pixel-based activity terms, with thresholds set from fifth percentiles of normal training windows.Event count and active-pixel fraction are computed by summing the voxel over time and polarity.Source: C. Learned and Activity-Supported… · p. 3
- An empty event window sets both injection weights to zero, so the updated pose equals the prediction.The weights multiply predicted candidate scales by deterministic activity support.Source: C. Learned and Activity-Supported… · p. 3
- The corrected pose becomes the next prediction history together with the preceding estimated pose.Source: C. Learned and Activity-Supported… · p. 3
D. Implementation and Training
The implementation encodes overlapping polarity-separated event windows and predicts rotation, image-plane translation, and correction scales. Training freezes candidate predictions before fitting the gate and evaluates self-fed histories only at test time.
D. Implementation and Training
- The event representation uses 200-ms windows at 100-ms intervals, five temporal bins per polarity, and 45 × 80 resolution.A ResNet-18-style encoder outputs 512 features.Source: D. Implementation and Training · p. 3
- The candidate loss applies Smooth-L1 to componentwise-normalized rotation and image-plane residuals.Training standard deviations normalize the residuals, with floors of 10^-4 rad and 10^-5.Source: D. Implementation and Training · p. 3
- The shared feature supplies rotation and image-plane candidates as well as the Scale Gate.The gate combines an event-feature branch with a scalar branch in a 160 → 128 → 2 sigmoid head and has 87,892 parameters.Source: D. Implementation and Training · pp. 3–4, Figure 3 · p. 4
- Candidate and gate training use AdamW for up to 30 epochs, with separate learning rates and batch sizes and three seeds.Development candidates are ranked by worst rotation/translation error ratio, then geometric mean, including an epoch-0 zero-correction candidate.Source: D. Implementation and Training · p. 4
- The selected candidate remains frozen during gate training, while event perturbations augment clean windows and gate checkpoints minimize masked scale-prediction loss.Reference pose histories are used in both training stages; self-fed histories are reserved for evaluation.Source: D. Implementation and Training · p. 4
IV. EXPERIMENTS
The experiment protocol centers on full-rollout self-fed tracking, with ablations and contextual measurements evaluated using the resulting tracker.
IV. EXPERIMENTS
- The main comparison follows self-fed tracking over the full rollout.Source: §IV · p. 4
- Update and injection ablations examine two design choices within the same tracking protocol.Source: §IV · p. 4
- Contextual comparisons and embedded measurements use the resulting tracker.Source: §IV · p. 4
A. Experimental Setup and Evaluation Protocol
Evaluation uses held-out FRESH physical-event trajectories, with method and checkpoint selection confined to training and development data. The protocol measures recursive pose errors and rollout failures, with endpoint gains and trajectory-cluster intervals reported for the four-trajectory cohort.
A. Experimental Setup and Evaluation Protocol
- The split contains 16 training, four development, and four held-out FRESH trajectories; parameter and checkpoint selection use only training and development partitions.The data are fixed-camera Prophesee EVK4 recordings with an aligned 800 × 720 central crop.Source: A. Experimental Setup and Evaluation… · p. 4
- 20 starts use two reference poses followed by 77 self-fed updates over 7.7 s; learned methods use three seeds, while the deterministic prior is evaluated once per start.Starts and target timestamps are shared across methods, and trajectories receive equal weight in aggregation.Source: A. Experimental Setup and Evaluation… · p. 4
- Rotation geodesic error, translation error in millimeters, and time-normalized error area are reported; failures begin at the first missing, non-finite, or invalid pose output.The last valid pose is held, and subsequent target-reference errors remain included in endpoint, area, and curve statistics.Source: A. Experimental Setup and Evaluation… · p. 4
- Fig. 4 pairs mean errors over 0.1–7.7 s with endpoint gains relative to the kinematic prior for each physical trajectory and cohort-level bootstrap intervals.The intervals are descriptive 95% trajectory-cluster bootstrap intervals.Source: Figure 4 · p. 5, A. Experimental Setup and Evaluation… · p. 5
B. Recursive Tracking Performance
In self-fed tracking, OA-EI yields lower recursive pose errors than the kinematic prior and full candidate injection. Its endpoint gains are positive on all four physical trajectories.
B. Recursive Tracking Performance
- After common initialization, each tracker feeds back its own estimates; full injection and OA-EI use the same rotational and image-plane candidates, while the prior uses constant-twist prediction.Full injection uses unit weights; OA-EI adds learned scaling and activity support.Source: B. Recursive Tracking Performance · p. 5
- 15.73°/22.41 mm at 7.7 s are OA-EI’s rotation and translation errors, reductions of 62.2%/60.9% from the prior; interval-wide error areas fall by 58.5%/55.8%.OA-EI has the lowest errors at all reported horizons.Source: B. Recursive Tracking Performance · p. 5
- 4/20 prior rollouts and 5/60 full-injection runs fail, compared with 1/60 for OA-EI; OA-EI endpoint gains are positive on all four trajectories.Failures are counted over the remaining rollout duration after the first invalid pose output.Source: B. Recursive Tracking Performance · p. 5
C. Matched State-Update Ablation
Matched comparisons favor OA-EI’s rotation-and-bearing update for overall pose accuracy, while combined scaling and activity support produce the lowest errors across tested event-retention levels. In the contextual pipeline comparison, OA-EI outperforms the best-performing baseline, with differing inputs and temporal mechanisms limiting direct comparison.
C. Matched State-Update Ablation
- OA-EI’s R+q interface has the lowest rotation, total translation, lateral error, and failure rate; adding depth correction lowers axial error to 6.71 mm from 7.26 mm but raises total translation error to 30.42 mm from 22.41 mm.The comparison includes R + XYZ residuals, R + q + z, and unified SE(3) updates, each with a separate scale fit under matched protocols.Source: C. Matched State-Update Ablation · pp. 5–6, Table II · p. 6
D. Injection Scaling and Sparse-Event Robustness
- Combining learned scaling and activity support gives the lowest errors at every tested event-retention level; at 1% retention OA-EI reaches 27.28°/38.64 mm while full injection exceeds prior errors.The ablation also reports 17/60 failures for full injection and 3/60 for OA-EI.Source: D. Injection Scaling and Sparse-Event… · p. 6, Table III · p. 6
E. Contextual Comparison with Existing Pipelines
- At 7.7 s, OA-EI records 15.73°/22.41 mm and 1/60 failures, versus 22.68°/34.37 mm and 4/60 for cascaded pose, the best-performing baseline.This is a contextual system-level comparison because the pipelines differ in input and temporal mechanisms; baselines use development-selected checkpoints and the shared held-out cohort and failure handling.Source: E. Contextual Comparison with Existing… · p. 6, Table IV · p. 6
F. Embedded Processing Cost
The measured in-memory update fits within the 100-ms budget on Jetson Orin Nano, while the conclusion reports lower recursive pose errors and failure rates for OA-EI. The evaluation remains limited to fixed-camera laboratory sequences initialized from two reference poses.
F. Embedded Processing Cost
- 44.56 ms median and 58.77 ms P95 in-memory update latency on Orin Nano both fall within the 100-ms budget for the measured processing segment.Timing includes voxelization, propagation, candidate/scale inference, and pose composition, but excludes sensor acquisition and online queue/slicing.Source: F. Embedded Processing Cost · p. 6, Table V · p. 6
V. CONCLUSION
- At 7.7 seconds, rotation and translation errors fall by 62.2% and 60.9% relative to kinematic prediction, while failure rates decrease from 20.0% to 1.67%.The state-update comparison favors rotation–bearing correction for overall pose accuracy, and combined Scale Gate and activity support outperform full injection at full, 10%, and 1% event retention.Source: §V · p. 6
- Evaluation covers fixed-camera laboratory sequences initialized with two reference poses, followed by recursive estimates; estimator-derived initialization and independently captured sequences remain unevaluated.Complete sensor-to-output latency is also left for future deployment evaluation.Source: §V · p. 7