Noise-Aware Learning for Traffic Safety Sensing: Propagation of Simulated Tracking Errors Through Surrogate Safety Measures and Denoising Models

Camera-derived traffic-conflict counts depend on trajectories produced by detection-and-tracking pipelines whose errors are rarely quantified. We measure that dependence with a controlled propagation study: seven types of simulated tracking error (independent and distance-dependent jitter, correlated drift, systematic offset, heading error, missed detections and track fragmentation), together with three held-out variants, are applied to drone-recorded reference trajectories from three signalized intersections with mixed traffic, a plausibility-filtered conflict screen based on time-to-collision (TTC) and post-encroachment time (PET) is rerun, and every resulting conflict is matched to the clean reference by road-user identity. The uncleaned condition is evaluated both for raw differentiated detections and for the output of a causal constant-velocity tracker, and four position-cleaning methods are compared, Savitzky–Golay smoothing, a Kalman–RTS smoother, and two learned denoisers of identical architecture, one trained on Gaussian jitter only and one on the matched distribution of the five position and detection error families, together with a gated hybrid that routes each track among them by its estimated noise level; all methods are evaluated both on raw detections and in the operational chain in which they follow the tracker. An analytical error model for constant-velocity TTC describes the onset mechanism: velocity errors from differentiation turn close parallel pairs into projected collision courses. Under tracker-like processing at 0.3 m mean position error, TTC precision pooled over road-user groups is 0.17 (0.09 for raw differentiated positions) against 0.66 for PET; errors that corrupt velocity bias TTC toward greater apparent severity, whereas constant offsets do not. The matched-distribution denoiser achieves the greatest improvement under position error (TTC precision 0.62 at 0.3 m on raw detections and 0.53 after the tracker, against 0.21–0.26 for the classical smoothers after the tracker) and generalizes to heavy-tailed jitter but adds nothing for missed detections; every cleaning method removes the TTC severity bias; the jitter-only model degrades detection under missed detections; the Kalman–RTS smoother is the most conservative method and the best under missed detections when applied to raw detections; and the gated hybrid removes the learned models’ loss on clean data when applied to raw detections but not to tracker output, whose filtering defeats the second-difference noise estimate. Fragmentation levels are compared with track-break rates observed at an instrumented intersection in New York City (COSMOS testbed). All results are conditional on the stated conflict definition, error models and thresholds, and are reported with pair-level confidence intervals.