Dataset

The emotion recognition component evaluates facial expressions and behavioral cues to infer potential intent and emotional states associated with suspicious activities. Data were collected through a combination of publicly available datasets, controlled experimental scenarios, and simulated radar measurements designed to represent realistic surveillance conditions. Each sensing modality produces a normalized threat score ranging from 0 to 1.

These scores are subsequently processed through a fuzzy logic-based decision-level fusion engine that generates a final threat assessment categorized into threat levels such as SAFE, SUSPICIOUS, and DANGEROUS. The dataset includes raw modality outputs, normalized threat scores, fusion results, evaluation metrics, graphical outputs, and supporting configuration files. The results demonstrate that multimodal fusion improves overall detection reliability, increasing ROC AUC from approximately 0.89 to 0.96 while reducing false alarm rates by approximately 17%.

The data further illustrate how complementary sensing information can resolve ambiguity that may exist within individual modalities. Researchers may use this dataset to investigate multimodal sensor fusion techniques, fuzzy inference systems, threat assessment algorithms, surveillance analytics, explainable artificial intelligence, and decision-support systems. The provided files enable replication of the fusion process, evaluation of individual modality contributions, and comparison with alternative fusion strategies.