Structural Verification for Occlusion-robust RGB-thermal Human Detection in Defense Surveillance via Part-graph Consistency and Evidential Fusion

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The partial occlusion is a major failure case of RGB-thermal (RGB-T) human detection in defense surveillance, where most end-to-end detectors generate the confidence score for the human detection in the form of a bounding box, but do not pay attention to whether the visible body parts configuration is anatomically plausible, leading to unreliable scores when such structural reasoning is most required.To fill this gap, we propose the Part-Graph Consistency Framework (PGCF) comprising three modules: Kinematic Constraint Likelihood (KCL), Graph Transformer Consistency (GTC) and Dirichlet evidential fusion (DEV).Geometric plausibility of detected part displacement is scored by KCL, global relational coherence is enforced by GTC, and the combined structural evidence is transformed to calibrated verification probabilities with explicit uncertainty estimates by DEV.Evaluated through five-fold cross-validation on CMSVD-a six-part annotated multispectral benchmark of 27,114 instances derived from KAIST and LLVIP-PGCF achieves SVA = 0.9049 ± 0.0052, SFPRR = 0.9967 ± 0.0010, ECE = 0.0283 ± 0.0011, and Brier = 0.0150 ± 0.0007, while preserving mAP@0.5 = 0.9747.Ablation shows that the contributions of each component are independent and that the most significant contribution is for severe occlusion.A detector-compatibility experiment on LLVIP with DAMSDet [29] yields preliminary evidence that PGCF can be integrated into an unseen detector architecture, and without loss of detection accuracy, without requiring fine-tuning.