Validated Adaptation for Aerial Crowd Monitoring at Mass Gathering Scale: A Deployment Protocol, a Severity Law, and a Diagnostic for Label-Free Drone Crowd Counting, Toward the FIFA World Cup 2034 (Saudi Arabia)

AlAnoud AllGhayth; AlJawharh AlOtaibi; Jude AlSubaie

Source-linked AI summary

Validated Adaptation for Aerial Crowd Monitoring at Mass Gathering Scale: A Deployment Protocol, a Severity Law, and a Diagnostic for Label-Free Drone Crowd Counting, Toward the FIFA World Cup 2034 (Saudi Arabia)

AlAnoud AllGhayth, AlJawharh AlOtaibi, Jude AlSubaie

arXiv:2608.17625v1Aug 18, 2026cs.AI

TL;DR

Aerial crowd counting for Saudi Arabia’s mass gatherings must adapt without labels to unfamiliar footage and warn of dangerous inflow before a crush forms. This paper validates a deployment protocol combining label-free adaptation, severity and stability analysis, and flux-based risk monitoring, recovering 30–49% of shift-induced error while repairing dense-scene undercounting.

Problem
Aerial crowd monitoring needs label-free counting that remains reliable on unfamiliar footage and detects dangerous crowd build-up early enough for intervention.
Method
The paper validates label-free adaptation with a conservation residual, flux-based congestion indicator, severity law, stability budget, and shift-gate evaluation.
Results
30–49% of shift-induced error was eliminated across four corruptions and five severities, while adaptation repaired dense-scene undercounting in the full-resolution corpus.
Takeaways & Limitations
The evidence supports unconditional adaptation with tail monitoring and a deployment protocol, while flux monitoring currently serves as congestion-ranking support rather than an absolute crush alarm.
Takeaways & Limitations
Absolute crush thresholds require meters-per-pixel calibration, so the flux module remains in ranking mode until a larger precision–recall and lead-time campaign is completed.

Abstract

from arXiv · show

Saudi Arabia will host the 2034 FIFA World Cup and already operates crowd management at Hajj scale. Drone-based counting must hold accuracy on footage unlike anything in its training corpus, without labels, and must warn of dangerous inflow before a crush forms. We deliver a validated answer built on 525 controlled runs, a full-resolution corpus study, five falsification ablations, and a five-condition safety-interlock evaluation. Label-free adaptation recovers 31-49% of shift-induced error across four corruptions and five severities, with the strongest method gaining 41.8 MAE over the frozen source (95% CI [34.1, 49.6], p=7.5x10^-10, d=2.52). We establish a severity law separating methods with a constant absolute margin from the one whose margin grows, and a stability budget identifying which configuration is safe to fly. On a full-resolution corpus carrying a genuine +48 MAE aerial gap (source retrained to 14.6 validation MAE, a 34% improvement), adaptation repairs the dense-scene undercounting that would otherwise under-report a forming crush, and the flux-based risk module fires on real congestion episodes in 2 of 6 full-length clips. We localise the recoverable error: in a regime built to favor a physics-informed conservation prior (300-frame clips at 200ms spacing, five times wider than standard), the adaptation signal is normalisation-driven, not flow-driven; the continuity residual is invariant to the proportional counting errors domain shift produces, confirmed by four on/off ablations correlated at r=0.999 and a 40% input corruption moving accuracy by only 0.05 MAE. A label-free shift gate shows shift magnitude and accuracy damage are rank-independent (Spearman rho=0.20; rho=-0.60 among genuine shifts), quantifying the 58% of headroom a magnitude gate forgoes. We establish unconditional adaptation with tail monitoring as policy, closing with a six-point protocol.

1 Introduction

This paper addresses label-free aerial crowd monitoring for Saudi Arabia’s FIFA World Cup 2034 and Hajj-scale gatherings, where domain shift and absent live ground truth can delay warnings of dangerous density build-up. It evaluates conservation-based test-time adaptation, severity behavior, transfer, diagnostics, and deployment safety on controlled corruptions and full-resolution footage.

Findings: 40–46% recovery of shift-induced error at reference severity and 30–49% across five severity levels establish effective adaptation.The evaluation uses DroneCrowd controlled corruptions and a five-level severity sweep.
Findings: The benefit does not degrade as corruption worsens, but the combined method incurs a stability cost with worst runs at low severity.The paper reports a per-method severity law and evaluates worst-run behavior rather than averages alone.
Findings: On full-resolution transfer, adaptation removes dense-scene undercounting, while the flux indicator fires on real congestion episodes.A full-resolution retrain improves validation MAE from 22.3 to 14.6, and the flux indicator is evaluated on real footage.
Findings: The conservation prior does not outperform entropy minimisation, including the wide-spacing regime designed to favor it, localizing recoverable error to normalisation statistics.The paper explains this result with an invariance argument and component-removing ablations.
Deployment implication: A label-free shift score poorly gates adaptation because shift magnitude does not track accuracy damage, supporting unconditional adaptation with worst-run tail monitoring.The system must adapt without live ground-truth counts while monitoring conditions where dangerous build-up can precede intervention.

2 Related Work

This section situates the study in label-free test-time adaptation, aerial crowd counting, physics-informed conservation priors, and shift detection. It adopts established counting infrastructure while framing conservation and shift magnitude as questions requiring task-specific validation.

Test-time adaptation: Test-time adaptation commonly intervenes in normalisation layers without target labels, using batch-statistic recomputation or entropy minimisation while convolutional weights remain frozen.AdaBN recomputes batch-normalisation statistics without gradient steps; TENT minimises prediction entropy through BN affine parameters.
Crowd counting: Density-map regression with dilated convolutions remains standard for congested scenes, and the study retains CSRNet unchanged to isolate the adaptation objective.DroneCrowd supplies the study corpus, representing the targeted mass-gathering setting through its scale, altitude range, and dense aerial viewpoints.
Physics-informed priors: Physics-informed learning constrains network outputs with laws; for counting, population conservation links regional count changes to boundary flux estimated from a displacement field.The section identifies conservation residuals as carrying useful information only when the conservation law genuinely constrains the solution.
Shift detection: Shift detection must support safety gating by ranking accuracy cost, not merely detecting whether distribution shift occurred.The study therefore treats shift magnitude as potentially unsuitable for deciding when to adapt.

3 Method

The method combines frozen-weight CSRNet counting with test-time batch-normalisation adaptation, a label-free conservation objective, an inward-flux congestion indicator, and a shift-triggered intervention gate. These components isolate adaptation losses while supporting both error diagnosis and early congestion monitoring.

Counting model: CSRNet maps each frame to a density map whose integral over region Ω yields the predicted count C_t(Ω).
Test-time adaptation: Only batch-normalisation parameters update on the test stream; convolutional weights remain frozen across Source, AdaBN, TENT, Ours, and TENT+Ours.This fixed-capacity design isolates the effect of each objective.
Physics-informed objective: The conservation residual uses pixel displacement from a frozen pretrained RAFT network and is minimized alone or alongside entropy loss as L = L_ent + λL_phys.The residual provides a label-free error signal when predictions violate continuity.
Congestion monitoring: The inward-flux signal Φ_t(Ω) identifies sustained positive inflow as a relative congestion-onset indicator before dangerous density develops.Absolute crush thresholds require people/m^2 density calibration and meters-per-pixel ground-plane scale.
Intervention policy: A shift gate computes scalar score s from incoming versus clean-data batch-normalisation statistics and adapts only when s > τ, with τ = 2s_clean.Section 8 evaluates this as a deployment decision policy against alternative policies.

4 Experimental Protocol

The experimental protocol combines a controlled corruption benchmark, a full-resolution real-motion corpus, and a safeguard-policy evaluation. Results are reported within track on fixed base models, with label-free adaptation updating only BN parameters and paired statistical analyses.

Track A: controlled corruption benchmark: n = 750 frames are evaluated under four fixed-content aerial corruptions: Gaussian noise, motion blur, low light, and JPEG compression.Each corruption is measured against a clean reference to isolate robustness from scene variability.
Track B: full-resolution corpus with real inter-frame motion: Validation MAE 14.6 is obtained by retraining the source model at full resolution, improving from 22.3 on the 11 GB corpus.Track B also uses real inter-frame motion and rebuilt pair sampling.
Metrics and analysis: MAE and RMSE are compared with paired t-tests, 95% confidence intervals, paired Cohen’s dz, Wilcoxon tests, and Holm–Bonferroni correction.Because replicates share seeds across methods, comparisons are paired; deterministic Source comparisons use one-sample tests.
Reporting discipline on base models: Track A uses a source model at validation MAE 26.1, while Track B uses the full-resolution retrain at 14.6; quantities are never pooled across tracks.Comparisons remain within-track on a single fixed base model, while the tracks are designed to converge on conclusions rather than absolute error levels.

5 Validated: Adaptation Efficacy and the Severity Law

Validated adaptation substantially reduces domain-shift error, with AdaBN gaining 41.8 MAE over the frozen Source model. Across severities, entropy-only adaptation preserves a near-constant absolute margin, while deployment selects BN realignment with entropy minimisation and excludes the combined objective because of tail instability.

Adaptation efficacy: 41.8 MAE is AdaBN’s gain over Source across 20 shifted replicates, with 95% CI [34.1, 49.6], p = 7.5 × 10−10, and Cohen’s d = 2.52.At reference severity, Source reaches 96.1 MAE, while adaptive methods reach 52–58 MAE, reducing shift-induced error by 40–45%.
Severity law: 35.7 MAE and 35.3 MAE are the protective margins recovered by entropy-only adaptation at severities 1 and 5, respectively.Source error rises from 73.6 to 112.0 MAE across severities 1–5, whereas severity-5 TENT reaches 76.7 and TENT+Ours 66.1.
Stability budget: 20.6% is the combined objective’s mean across-replicate CV, versus 11.3% for TENT and 10.9% for Ours.TENT+Ours peaks at 70.8% in one cell and reaches a worst individual run of 113.5 MAE, compared with 91.8 for TENT.
Full-corpus validation: Adaptation cuts counting error in every full-corpus density band and moves bias toward zero, with the largest absolute correction in the densest scenes.The densest-scene regime is identified as crush-relevant because negative bias represents undercounting.
Deployment decision: BN realignment with entropy minimisation is validated for deployment, while the combined objective is held back from flight because of tail behaviour despite superior mean performance.The selected adaptation stage must operate within the stability budget of Section 10.

6 Validated: Full-Corpus Transfer · 7 Determined: Where the Adaptation Signal Comes From

Full-resolution transfer validates label-free adaptation on a genuine aerial gap, especially dense-scene undercounting, while ablations show the recoverable signal is normalization- and confidence-driven rather than flow-driven. The resulting mechanism specifies when conservation should be retested and how adaptation should be deployed under appearance shift.

6 Validated: Full-Corpus Transfer: 14.6 against 22.3, a 34% improvement, was achieved by ingesting the full 11 GB release, converting native trajectory annotations, and retraining at full resolution.This stronger source model raises the baseline for downstream adaptation claims.
6 Validated: Full-Corpus Transfer: +48 MAE degradation on target scenes was largely removed by adaptation, which reduced densest-scene MAE from 194.7 to 98.5 while addressing dangerous undercounting.The source error was dominated by systematic dense-scene undercounting that could under-report a forming crush.
6 Validated: Full-Corpus Transfer: 200 ms spacing and full-length clips supplied the wide-frame and operational conditions needed to evaluate the mechanism and make the risk module measurable.The full-corpus track used full-resolution retraining, real motion, and clips suitable for downstream safety analysis.
7 Determined: Where the Adaptation Signal Comes From: 1.71 MAE separated entropy minimisation from the conservation objective across 100 paired severity-sweep runs (95% CI [1.50, 1.93]; paired p = 3.9 × 10−29; dz = 1.60).The differences remained consistent across Gaussian noise, JPEG, low light, and motion blur, surviving Holm correction.
7 Determined: Where the Adaptation Signal Comes From: 52.41 versus 52.13 on clean full-corpus data yielded ∆= +0.28 (95% CI [−0.28, 0.84], p = 0.24), while low light yielded ∆= −0.05 when Lphys was toggled.The four-regime toggle traces followed the same trajectory run for run, isolating little usable contribution from the physics term.
7 Determined: Where the Adaptation Signal Comes From: 40% optical-flow input noise moved MAE by only 0.05 (43.86 →43.91), indicating that the flow-based conservation input was not materially being used.Input corruption complements toggling by testing whether an auxiliary objective’s input affects output.
7 Determined: Where the Adaptation Signal Comes From: The continuity residual is invariant to proportional counting errors from appearance shift, so adaptation capacity should target normalisation statistics and prediction confidence rather than flow-based conservation.This explains why appearance corruptions can preserve frame-to-frame count balance despite damaging absolute counts.
7 Determined: Where the Adaptation Signal Comes From: Conservation should be decisively retested under occlusion, frame-boundary entry and exit, or tracking-scale flows through gates and concourses, where count balance itself breaks.The shift-gated policy uses τ = 2sclean = 0.0022, while stability analysis identifies TENT+Ours as having the widest extreme tail, near 113 MAE.

8 Determined: Shift Magnitude Does Not Predict Harm

Shift magnitude did not predict adaptation harm: moderate corruptions caused severe accuracy damage despite barely moving batch-normalisation statistics. Because adaptation improved accuracy in all five tested conditions, unconditional adaptation outperformed magnitude-gated policies.

Gate extremes: 102.3 →67.1 and 81.2 →46.4 MAE under the two strongest shifts showed that the gate correctly fired when statistical displacement was large.The gate abstained on clean data, where only 6% was available, avoiding adaptation spending when none was warranted.
Moderate shifts: 93.4 and 87.5 source MAE fell to 42.2 and 44.3 under adaptation for motion blur and JPEG, recovering 55% and 49% of error.Their shift scores, s = 0.00207 and 0.00135, were both below τ = 0.0022, so the magnitude gate would miss them.
Interpretation: The study identifies displacement and damage as different quantities, making shift-magnitude gating an unreliable safety interlock for unsupervised deployment.A gate remains desirable only if evaluated regimes include cases where adaptation genuinely degrades accuracy or the score predicts harm rather than statistical distance.
Policy result: 80.3 mean MAE fell to 46.9 under unconditional adaptation, which captured 100% of available headroom versus 42% for magnitude gating.The oracle policy matched unconditional adaptation because adaptation was better in all five conditions, including clean data.

9 Risk Alerting on Full-Length Clips

On six full-length clips, the flux-based alert recovered every annotated danger frame in one of two genuine danger scenes, with a 4.4 s mean lead but low precision, while missing the other episode. The module therefore ranks congestion onset provisionally; absolute crush thresholds await density calibration and broader precision–recall evaluation.

Full-length clip evaluation: 2 of 6 target scenes contained genuine danger episodes, and the flux indicator recovered every annotated danger frame in one scene but did not trigger in the other.The missed episode was a false negative.
Full-length clip evaluation: Recall 1.00 and a mean lead of 4.4 s before onset were achieved in the detected episode, at the cost of precision 0.23.The indicator also fired frequently before the annotated danger period.
Validation limits: With two positive episodes, the module is an established response, but larger positive sets are needed for precision–recall and lead-time characterisation.The protocol schedules this data requirement.
Validation limits: Absolute crush thresholds require density calibration in people/m^2 from meters-per-pixel scale to the ground plane; until then, the module ranks congestion onset rather than asserting absolute risk.This limitation applies before the planned calibration campaign.

10 Deployment Protocol

The deployment protocol resolves the study into six actionable rules: adapt unconditionally, prioritize normalization and confidence, monitor instability tails, calibrate alarm use, and validate venue-specific safety thresholds before deployment.

Deployment Protocol: 100% of available headroom is captured by unconditional adaptation, versus 42% for a magnitude-based gate.Adaptation was better in every tested condition and is the derived-optimal policy.
Deployment Protocol: BN realignment plus entropy minimisation is the single-objective adaptation stage carrying validated accuracy and the tighter stability envelope.The protocol specifies one adaptation objective rather than stacking multiple adaptation mechanisms.
Deployment Protocol: Adaptation capacity should target normalisation and confidence, not flow-based conservation, because the continuity residual is invariant to proportional counting error.Proportional counting error is the error appearance produced by domain shift.
Deployment Protocol: CV ≤∼12% and worst-run degradation bounded relative to the median define the tail budget reported alongside MAE.Instability at low shift is inherited by stacked components, so it must be monitored rather than assumed away.
Deployment Protocol: The flux alarm should operate in ranking mode as an operator aid, with automatic triggering gated behind metric calibration.This separates operational ranking support from an automatic safety interlock.
Deployment Protocol: Calibration before the venue must combine meters-per-pixel scale, congested ingress/egress footage, and labelled corruption sweeps to unlock validated crush thresholds, lead-time curves, and interlocks.Each output has a defined acceptance criterion and is scoped by the study.

11 Scope and Operating Envelope

The study defines a bounded operating envelope validated by two complementary tracks, with within-track comparisons and a single fixed base model. Its safety components support response and ranking, but not yet absolute alarm claims or broader architecture generalization.

Operating envelope: The deployment result is reported within a precisely stated operating envelope rather than as an unrestricted claim.The authors frame the envelope as the boundary within which the result is known to hold.
Validation tracks: Two complementary tracks corroborate shared conclusions across controlled synthetic corruptions and a genuine domain gap on independently retrained full-resolution data.Track A enables exact causal attribution by varying only corruption, while Track B addresses realism; their agreement is presented as internal validation before event data.
Comparison design: Comparisons are made only within each track, against one fixed base model, because the tracks use different absolute error scales.Recurring conclusions across two independently trained backbones and error scales are therefore treated as robust rather than incidental.
Architecture scope: The reported architecture scope is CSRNet with RAFT, while the continuity-residual diagnosis is argued as architecture-independent and input-corruption ablation excludes an estimator-specific explanation.Confirmation on a second density parameterisation is scheduled, not presented as necessary to resolve the mechanism.
Safety operating mode: The flux indicator establishes response to genuine congestion and is therefore presented in ranking mode, while the shift gate’s damage finding is threshold-independent under the evaluated rule.A lead-time calibration curve remains to be completed, so the flux component is not reported as an absolute alarm.

12 Conclusion

The study defines when label-free test-time adaptation should be used for aerial crowd-monitoring safety and supports deployment through controlled validation, diagnostics, and a reproducible protocol. It localises the adaptation signal to normalization under appearance shift and provides measurement tools intended for operational footage, including the 2034 FIFA World Cup in Saudi Arabia.

Conclusion: Across 525 controlled runs and a full-resolution corpus study, adaptation eliminates 30–49% of shift-induced error across four corruptions and five severities.The study also defines a severity law describing how each method’s protective margin behaves as conditions deteriorate.
Conclusion: Under appearance shift, recoverable error is normalization-borne, while a flow-based conservation residual remains invariant to the resulting proportional counting error.This finding was tested across two corpora, two frame rates, and five ablations, including a regime designed to favor the conservation prior.
Conclusion: The work delivers a deployment protocol, calibration campaign with defined acceptance criteria, and measurement apparatus applicable unchanged to footage including the 2034 FIFA World Cup in Saudi Arabia.The apparatus includes a paired-seed design, two falsification ablations, and policy-level safety-gate evaluation.
Conclusion: Every reported number derives from released run tables and analysis scripts covering the benchmark, severity sweep, conservation ablations, flow-corruption sweep, and safeguard evaluation.The released materials compute every reported interval and p-value.