Multimodal deep learning framework for real-time detection of gender-targeted criminal activities using surveillance videos in process

Debilitation effects of increasing women's crimes such as stalking and assault harassment and snatchings on the women's need for surveillance systems in the detection process, which should be real time and context aware. Based on existing approaches to anomaly detection that make use of video, generalizations that lack gender-specific contextual modeling, have biased training datasets, and insubstantial intent-driven interaction modeling can be identified in most cases. Additionally most models are not generalizable across diverse video sources and do not have an uncertainty handling mechanism for high-risk decisions. To this end, the current study present a well-detailed deep learning framework for detecting criminal activities against women using multi-source video surveillance data. A dataset of 637 such videos curated with contributions from social media sources, crime series, and police records is classified under five behavior-specific classes. The five novel comprehensions applied in the proposed system emanate from: (i) Spatio-Temporal Graph Attention Transformer (ST-GAT)-modeled human-object interactions and collective behavior; (ii) an Adversarial Domain-Invariant Video Encoder (ADIVE) without platform-specific feature biases; (iii) Self-Supervised Scene Graph Contrastive Learning (SSCGL) about semantic anomalies in relational contexts and (iv) a Multimodal Cross-Attention Fusion Network (MCAFNet) within synchronized audio cues to enhance detection of verbal abuse or threats and a Bayesian Uncertainty-Aware Decision Calibration (BUADC) component, which includes confidence-aware predictions for safe deployments. These systems improve context comprehension, cross-domain generalizability, and false alarm reduction. Model test accuracy is 87%. Critical classes like "Chasing, Stalking, Snatching" are 0.93 precise. This methodology provides scalable, interpretable, high-confidence monitoring for women. It has global law enforcement potential.