Budget-Aligned Epistemic Uncertainty for Onboard UAV Trajectory Prediction via Regression-Adapted Deep Deterministic Uncertainty

Onboard UAV trajectory prediction for safetycritical missions requires not only accurate forecasts but also deployment-oriented epistemic uncertainty under strict millisecond-level inference budgets. Existing evaluations are often budget-misaligned, comparing single-forward predictors with multi-forward baselines without normalizing inference cost. In this paper, a budget-aligned evaluation protocol is established based on the forward-pass budget, floating-point operations (FLOPs), and CPU-proxy latency. Outage-segment trajectorylevel survival is introduced to assess system-level safety. Under the strict single-forward constraint, deep deterministic uncertainty (DDU) is adapted to time-series regression by stabilizing the feature space with spectral normalization, modeling feature density via a Gaussian mixture model (GMM), and mapping density scores to continuous epistemic variance through isotonic calibration. Stress tests on a physics-intensity grid demonstrate that the proposed approach retains safety under severe runtime drift. In the HighDyn_LongOut scenario, a trajectory-level survival probability of 0.92 is achieved, while a five-member deep ensemble baseline yields 0.45. The proposed post-processing introduces only about 8.5% overhead and provides approximately 4.6 times speedup relative to the ensemble baseline under CPUproxy evaluation, supporting onboard-feasible epistemic uncertainty estimation within a millisecond-level budget under the evaluated simulation setting.