Person detection · temporal tracking · false-alarm trade-offs · track-fragment association
A reproducible study of a practical railway-vision question:
Can temporal processing recover people missed by a frame-level detector without producing an unacceptable number of false alarms and incorrect cross-person associations?
Dataset: OSDaR23 · Research: completed · Final stage: B7-SCF
Scientific decision: FAIL — predefined hypothesis not confirmed
Computation: FROZEN_AFTER_B7 · Sealed test access: 0
One frame can miss a person. Temporal tracking can recover some of these misses, but may also create additional false tracks and incorrect associations. This project tests whether recovery can be improved without paying an unacceptable safety-related cost.
flowchart LR
A[Railway video] --> B[Person detector]
B --> C[Detected person]
B --> D[Missed person]
B --> E[False alarm]
D --> F[Temporal tracking]
F --> G[Recover missed observations]
G --> H{Trade-off}
H --> I[More recovered persons]
H --> J[More false tracks or wrong associations]
J --> K[Verification and fragment association]
K --> L[Can recovery improve safely?]
Research: COMPLETED
Final computational stage: B7-SCF
Scientific gate: SCIENTIFIC_FAIL
Operational gate: OPERATIONAL_FAIL
Computation state: FROZEN_AFTER_B7
B8: PROHIBITED BY FROZEN PROTOCOL
Test: SEALED
Test access count: 0
SCIENTIFIC_FAIL means that the prospectively defined hypothesis did not pass
all scientific gates. It does not mean that the software failed or that the
experiment is invalid. OPERATIONAL_FAIL is a separate finding: the complete
set of predefined operational requirements was not met. Both are final,
reproducible negative results.
After B7, the computation was frozen. The same evidence base must not be used for B8, new threshold selection, weight changes, consensus tuning, or opening the sealed test. A future study would require an independent experimental basis.
The project studies whether missed person detections in independent railway video scenes can be reduced through temporal processing, tracking, and track-fragment association without causing an unacceptable increase in false alarms or cross-person merges.
This is one research project. TNormFilter is the historical name of an early
diagnostic code module. Operator Assistant is the internal name of a later
human-review experiment. B6 and B7 are sequential track-fragment association
stages; they are not separate products or research projects.
The repository is the public software and reproducibility record for the associated manuscript. It contains source code, frozen protocol definitions, aggregated results, threshold analyses, trace audits, and a non-autonomous review demonstrator. It is not a production or safety-certified system.
flowchart LR
A[OSDaR23 railway scenes] --> B[Person detection]
B --> C[Cross-scene generalization]
C --> D[Robustness and T-norm diagnostics]
D --> E[Threshold trade-off]
E --> F[Temporal tracking]
F --> G[Track verification]
G --> H[TrackFragment construction]
H --> I[B6 pairwise association]
I --> J[B7 global flow]
J --> K[B7 hard constraints]
K --> L[B7 consensus and SCF]
L --> M[Predefined gates]
M --> N[SCIENTIFIC_FAIL]
N --> O[FROZEN_AFTER_B7]
| Stage | Research question | Frozen result |
|---|---|---|
| Detection | Can people be detected reliably across independent railway scenes? | A baseline was obtained; a substantial cross-scene gap remained. |
| Pretraining / robustness | Do CrowdHuman pretraining and robustness methods close that gap? | The tested configurations did not fully close it. |
| T-norm diagnostics | Do T-norm features add stable information beyond detector and representation features? | No additional value was confirmed on the primary endpoint. |
| Threshold analysis | Can misses be recovered by lowering the frame-level operating threshold? | Recall increased together with false alarms. |
| Tracking | Does temporal continuity reduce misses? | FN decreased, but false alarms increased sharply. |
| Verification | Can false tracks be removed while keeping recovered observations? | The trade-off remained. |
| B6-FULL | Can a pairwise classifier associate track fragments reliably? | Many true fragment links were missed. |
| B7-FLOW | Does global path optimization improve recovery? | Recovery improved, but cross-person merges increased. |
| B7-HARD | Do hard constraints control aggressive flow linking? | Aggressiveness decreased, but the scientific gates were not met. |
| B7-CONSENSUS | Does perturbation consensus provide a stable acceptable solution? | Merges decreased, but useful links were nearly suppressed. |
| Final | Were all predefined gates met? | No: SCIENTIFIC_FAIL, OPERATIONAL_FAIL, FROZEN_AFTER_B7. |
Historical protocols and results remain in the repository for provenance. They are completed stages, not open tasks.
- Temporal processing recovered approximately
0.13Recall and reduced FN/frame by roughly17–19%, but increased false alarms by276–391%. - T-norm features did not improve the frozen V8b primary endpoint.
- B7-FLOW recovered more correct fragment links, but cross-person merges rose
to
0.40000. - B7-CONSENSUS reduced merges, but suppressed most useful links and improved
0/5scenes. No tested method passed every predefined gate.
The railway experiments use RGB material from OSDaR23. OSDaR23 is obtained separately under its own distribution conditions; raw images and annotations are not distributed in this repository. The project treats a grouped scene, not an adjacent frame or tile, as the independent unit for splits and scene statistics.
CrowdHuman was used only in designated pretraining experiments under its non-commercial research/education terms. CrowdHuman images, archives, derived datasets, crops, and checkpoints are not included in the repository or public evidence.
The final B6/B7 association benchmark has the following frozen scale:
Development scenes: 5
Frames: 479
Person observations: 4943
TrackFragments: 646
Reference PersonEpisodes: 48 (evaluation only)
The railway test contains five held-out grouped scenes. It remained sealed
through the final B7 decision. Test access count: 0.
The V8b primary comparison tested whether Product and Łukasiewicz consistency
features improved prediction of FN/frame over detector and standard
representation features:
U2 scene-macro MAE: 1.59549
U3 scene-macro MAE: 1.76798
Relative change: -10.81%
Scene wins: 5/15
The added T-norm features worsened the primary MAE in this setting.
Temporal processing consistently recovered some missed observations. Across
the studied variants, Recall increased by approximately 0.13 and FN/frame
fell by approximately 17–19%. The same variants increased false alarms by
approximately 276–391%, so the practical deployment gate failed. These
results document an operating-point trade-off, not a deployable safety system.
The B6 audit restored full link-level observability:
TP links: 44
FP links: 16
FN links: 750
True positive pairs: 794
Predicted positive pairs: 60
Model coefficients: 95
Scaler parameters: 95
B6 pseudo-pair rows: 36864
The final comparison uses frozen micro metrics and perturbation stability:
| Method | Micro F1 | Cross-person merge | Split recovery | Median ARI |
|---|---|---|---|---|
| B6-FULL | 0.10304 | 0.26667 | 0.05542 | 0.53446 |
| B7-FLOW | 0.19600 | 0.40000 | 0.11713 | 0.35724 |
| B7-HARD | 0.10514 | 0.27419 | 0.05668 | 0.61004 |
| B7-CONSENSUS | 0.03206 | 0.23529 | 0.01637 | 0.61780 |
Global flow increased Micro F1 from 0.10304 to 0.19600 and split recovery
from 0.05542 to 0.11713. This confirms that global path optimization can
recover additional correct links. It was not accepted because cross-person
merge increased from 0.26667 to 0.40000, while median perturbation ARI fell
from 0.53446 to 0.35724.
Hard constraints reduced the aggressiveness of flow, yielding Micro F1
0.10514, cross-person merge 0.27419, split recovery 0.05668, and median
ARI 0.61004. The combined improvement was insufficient for the frozen gate.
The primary consensus variant reduced cross-person merge to 0.23529 and
raised median perturbation ARI to 0.61780, but its Micro F1 fell to 0.03206
and split recovery to 0.01637. It improved 0/5 scenes. Its 10th-percentile
ARI was 0.46926, coverage was 0.90462, and the prospectively required
median ARI was at least 0.68446. Consensus was too conservative and
suppressed most useful links.
flowchart TD
A[B6-FULL<br/>F1 0.10304<br/>Merge 0.26667] --> B[B7-FLOW]
B --> C[Recovery improves<br/>F1 0.19600]
B --> D[Cross-person merge rises<br/>0.40000]
C --> E[B7-HARD and CONSENSUS]
D --> E
E --> F[Merge decreases]
E --> G[Correct links collapse<br/>F1 0.03206]
F --> H{All predefined gates passed?}
G --> H
H -->|No| I[SCIENTIFIC_FAIL]
I --> J[FROZEN_AFTER_B7]
J --> K[B8 prohibited on same evidence]
The prospectively registered B7 hypothesis was not confirmed. Global flow improved recovery but worsened cross-person association. Stability consensus reduced merges but nearly eliminated correct links. None of the variants simultaneously improved recovery, controlled cross-person merges, improved cluster stability, and passed all predefined gates.
Scientific decision: FAIL
Operational decision: FAIL
Compute status: FROZEN_AFTER_B7
B8 allowed: false
The final scientific result is therefore the documented boundary of this trade-off, not a claim that track association or railway hazard detection has been solved.
ИССЛЕДОВАНИЕ КОМПРОМИССА МЕЖДУ ПРОПУСКАМИ И ЛОЖНЫМИ ТРЕВОГАМИ ПРИ ОБНАРУЖЕНИИ ЛЮДЕЙ В ЖЕЛЕЗНОДОРОЖНЫХ ВИДЕОСЦЕНАХ
A STUDY OF THE TRADE-OFF BETWEEN MISSED DETECTIONS AND FALSE ALARMS IN RAILWAY PERSON DETECTION
Yuri V. Trofimov · Alexey N. Averkin · Alexey V. Shevchenko · Tatiana V. Kim · Alexander D. Lebedev
- Publication status: manuscript
- DOI: pending
- Public article URL: pending
- Research repository: github.com/lebedeffson/railway-vision-robustness
- Public evidence:
artifacts/operator_assistant_b7_public.zip
No DOI or publisher URL is assigned in the available final manuscript, and no placeholder DOI has been created.
Исследование выполнено в рамках государственного задания Министерства науки и высшего образования Российской Федерации, тема № 124112200072-2.
This research was carried out within the framework of the State Assignment of the Ministry of Science and Higher Education of the Russian Federation, topic No. 124112200072-2.
Several authors are affiliated with Dubna State University. This statement does not claim that the repository is legally owned by the university.
The final B7 provenance commit is
f25a751c52104b0353995d2206afbe92a4182ff2.
The computation lock records its immediate predecessor
484d5eebf1bdc6a02cfeb7f751e81443d881c2f3. Both remain reachable from
main; their history must not be squashed or rebased.
The tracked public bundle contains the frozen aggregated evidence, protocol locks, hierarchy, manifest, replay audit, and final decision. Verify it with Python's standard library:
git clone https://github.com/lebedeffson/railway-vision-robustness.git
cd railway-vision-robustness
python scripts/operator_assistant_b7/verify_public_evidence.py \
artifacts/operator_assistant_b7_public.zip --jsonExpected status:
manifest: PASS
b6_audit: PASS
b7_metrics: PASS
replay: PASS
cross_process_determinism: PASS
scientific_gate: FAIL
operational_gate: FAIL
test_status: SEALED
test_access_count: 0
The frozen B7 bundle records the original finalization suite (481 passed).
After adding repository-publication checks, the complete suite on the internal
experiment checkout is 484 passed; the scientific result files were not
changed. A clean public checkout reports 144 passed, 340 skipped: the skipped
integration tests require private frozen outputs or checkpoints that are
intentionally not distributed. Canonical B7 replay is included in the 144
passing public tests and is not skipped.
Full recomputation requires separately obtained OSDaR23 inputs and the frozen
derived B6 evidence identified by the hashes in
B7_PROTOCOL_LOCK.json.
The repository never downloads data or model weights implicitly.
git clone https://github.com/lebedeffson/railway-vision-robustness.git
cd railway-vision-robustness
python -m venv .venv
.venv/bin/pip install -r requirements-dev.txt
PYTHONPATH="$PWD:$PWD/scripts" .venv/bin/python \
scripts/operator_assistant_b7/run_b7.py prepare
PYTHONPATH="$PWD:$PWD/scripts" .venv/bin/python \
scripts/operator_assistant_b7/run_b7.py run
PYTHONPATH="$PWD:$PWD/scripts" .venv/bin/python \
scripts/operator_assistant_b7/finalize_b7.py
.venv/bin/python -m pytest -qprepare validates and locks the already-frozen inputs. It does not authorize
new model training, threshold tuning, B8, or access to the sealed test.
- Bundle:
operator_assistant_b7_public.zip - SHA-256:
7f4892b170fbb4da9721077f201dbfeeafa5906bfd16bac8745b31b76c3dc09d - Sidecar:
operator_assistant_b7_public.zip.sha256 - Manifest:
MANIFEST.sha256inside the archive
The bundle includes B7_HIERARCHY.jsonl, the file referenced by the
cross-process determinism report. It also includes B6 audit files, B7 candidate
and selected edges, model coefficients, scene normalization, pseudo-pair
indices, per-scene/overall results, protocol locks, final decisions, and
figures.
The bundle excludes raw images, video, restricted annotations, model checkpoints, crops, feature tensors, the sealed test, local absolute paths, credentials, and internal project memory.
configs/ frozen experiment configurations
protocol/ immutable protocol locks and final decisions
scripts/ reproducibility, evaluation, audit, and packaging commands
src/ detector, tracking, verification, and review modules
tests/ regression, integrity, leakage, and release checks
artifacts/ compact public evidence release
reports/ technical reports from completed stages
article/ manuscript-support tooling, not the publication itself
README.md is the public scientific source of truth. AGENTS.md contains
development constraints, while .codex/ stores internal workflow notes.
- The association evaluation contains five development scenes, 48 reference person episodes, and 646 fragments; statistical power and visual diversity are limited.
- The results concern the studied railway RGB scenes and do not automatically transfer to other cameras, stations, routes, weather, or operating regimes.
- The detector used in parts of the development history was not fully detector-OOF for every scene.
- No operator or human-factors study was performed. Fewer cards or detections do not establish reduced cognitive workload.
- There is no safety certification, production-readiness claim, autonomous alarm output, or safety actuation.
- Temporal modes retain a high false-alarm rate.
- The sealed test was not opened; no test-performance claim is made.
- B8 and further tuning on this evidence base are prohibited by the frozen protocol.
Until a DOI is assigned, cite the manuscript without invented publication metadata:
Trofimov, Yuri V.; Averkin, Alexey N.; Shevchenko, Alexey V.;
Kim, Tatiana V.; Lebedev, Alexander D.
“Исследование компромисса между пропусками и ложными тревогами при
обнаружении людей в железнодорожных видеосценах”. Manuscript.
Software and evidence: Railway Vision Robustness,
https://github.com/lebedeffson/railway-vision-robustness
Machine-readable software citation metadata is available in
CITATION.cff. Add the DOI and publisher URL only after they
exist in an official public record.
License decision required from repository owner. No software or data license has been selected automatically. Third-party datasets and pretrained weights retain their own terms; repository visibility alone does not grant permission to reuse them.