Target-count alignment
Scene-level comparison of observed target-count error.
Study scope
HG-SMG-TC combines homography-guided endpoint micro-mode discovery, semantic approach consolidation, semantic maneuver graph construction, uncertainty-aware maneuver-count priors and prior-constrained model selection.
The evaluation is limited to five Bellevue scenes and uses a leakage-controlled split protocol. The method is compared against original HG-aware selection and trajectory-clustering baselines using independently generated polygon-rule reference labels.
Five Bellevue scenes
- bellevue_116th_ne12th
- bellevue_150th_newport
- bellevue_150th_eastgate
- bellevue_150th_se38th
- bellevue_ne8th
Method pipeline
Locked-test result summary
| Method family | Target error | NMI | Macro F1 | Outlier % |
|---|---|---|---|---|
| Original untargeted A0 | 3.2000 | 0.8155 | 0.6328 | 15.56 |
| Original HG-aware A1 | 2.5333 | 0.8376 | 0.6838 | 9.94 |
| HG-SMG-TC A5 | 2.1333 | 0.8386 | 0.6723 | 10.34 |
| Endpoint isotropic baseline | 3.4667 | 0.7099 | 0.5710 | 24.43 |
| Resampled trajectory baseline | 3.0000 | 0.6827 | 0.4752 | 25.02 |
The result should be read conservatively: HG-SMG-TC improves mean target-count alignment and slightly improves NMI relative to the original HG-aware method, while macro F1 and outlier percentage show trade-offs.
Evidence figures
Independent agreement metrics
ARI, NMI, purity and mapped F1 are interpreted as independent validation metrics.
Baseline comparison
Compact summary against endpoint and resampled-trajectory baselines.
SE38th limitation
Diagnostic evidence for semantic over-segmentation and residual trade-offs.
Data and reproducibility
GitHub contains code, configurations, frozen protocols, compact result tables, final evidence figures, reference-label documentation, homography quality summaries and reproducibility documentation. OneDrive is used for large redistributable artifacts such as videos, tracking outputs and processed trajectory data. The release does not claim that true raw per-frame YOLOv11 detector exports are available.