Table 3.

Ablation study results on UA-DETRAC and CityFlow datasets, analyzing the impact of removing key components—Relational Attention Mechanism, Risk-Oriented Scene Decoding, and Directional Local Risk. The full model achieves the highest performance across all metrics.

Model	UA-DETRAC Dataset				CityFlow Dataset
	Accuracy	Precision	F1 Score	AUC	Accuracy	Precision	F1 Score	AUC
w./o. Relational Attention Mechanism	90.17 ± 0.02	86.03 ± 0.03	87.55 ± 0.02	89.74 ± 0.02	88.33 ± 0.03	84.50 ± 0.02	85.61 ± 0.02	88.46 ± 0.03
w./o. Risk-Oriented Scene Decoding	91.22 ± 0.03	88.71 ± 0.02	88.45 ± 0.03	91.33 ± 0.03	89.66 ± 0.02	86.79 ± 0.03	87.24 ± 0.02	90.14 ± 0.02
w./o. Directional Local Risk	89.38 ± 0.03	85.90 ± 0.02	86.31 ± 0.03	88.82 ± 0.02	87.95 ± 0.02	84.18 ± 0.03	84.77 ± 0.02	87.91 ± 0.03
Ours	92.85 ±0.02	89.33 ±0.02	90.12 ±0.03	93.21 ±0.02	91.90 ±0.02	88.67 ±0.03	89.45 ±0.02	92.38 ±0.03