Scholarlink

<p>With the growing concern in urban environments, the existence of intelligent systems that detect women-at-risk crimes in real-time has become very crucial in process. The traditional passive surveillance systems heavily depended on monitoring and shallow pattern recognition, lacking skills to see the complexities of human interactions: simultaneous actions, and multi-view spatial context especially in dense or ambiguous situations. Existing efforts treat scenes like isolated incidents, ignoring the evolution over time of the threats themselves, interpersonal dynamics involved, and the issue of preserving privacy while monitoring movements between locations. In addressing these issues, this work proposes a full multi-stage anomaly detection framework with five brand new modules, each one targeting a certain dimension that facets into real criminal scenarios. The Spatio-Temporal Transformer Network with Attention-Guided Threat Mapping (STTN-AGTM) using multi-scale embeddings and a threat-based attention mechanism to localize and quantify spatial risk zones. The Multi-Instance Multi-Label Crime-Specific Latent Embedding (MIML-CLE) module allows for the detection of co-occurring micro-actions like aggression and proximity through fine action clustering. The Crime-Aware Graph Convolutional Network with Dynamic Node Contextualization (CA-GCN-DNC) detects group-based threats by modeling inter-person interactions as evolving graphs. The Hyperparameter Optimized Ensemble with Risk-Adaptive Learning (HOERAL) uses Bayesian optimization to adaptively weight outputs from different classifiers for improved generalization and confidence. Last but not least, Real-Time Federated Learning with Cross-Camera Collaboration (RTFL-CCC) should enable privacy-compliant tracking across cameras and collaborative inference of threats. This integrated architecture has shown very good accuracy (up to 95.8%) to detecting and localizing threats while gaining substantially on action co-occurrence, relational reasoning, and false-positive mitigation toward the process. The proposed system thus sets a strong, scalable, and ethically compliant paradigm for real-time anomaly detection in public safety surveillance settings.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Design of an integrated novel method for real-time anomaly detection against women using multimodal spatio-temporal, graph-based, and federated learning architectures

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Kanchan Ganesh Dhuri,
  • Sunita Patil

摘要

With the growing concern in urban environments, the existence of intelligent systems that detect women-at-risk crimes in real-time has become very crucial in process. The traditional passive surveillance systems heavily depended on monitoring and shallow pattern recognition, lacking skills to see the complexities of human interactions: simultaneous actions, and multi-view spatial context especially in dense or ambiguous situations. Existing efforts treat scenes like isolated incidents, ignoring the evolution over time of the threats themselves, interpersonal dynamics involved, and the issue of preserving privacy while monitoring movements between locations. In addressing these issues, this work proposes a full multi-stage anomaly detection framework with five brand new modules, each one targeting a certain dimension that facets into real criminal scenarios. The Spatio-Temporal Transformer Network with Attention-Guided Threat Mapping (STTN-AGTM) using multi-scale embeddings and a threat-based attention mechanism to localize and quantify spatial risk zones. The Multi-Instance Multi-Label Crime-Specific Latent Embedding (MIML-CLE) module allows for the detection of co-occurring micro-actions like aggression and proximity through fine action clustering. The Crime-Aware Graph Convolutional Network with Dynamic Node Contextualization (CA-GCN-DNC) detects group-based threats by modeling inter-person interactions as evolving graphs. The Hyperparameter Optimized Ensemble with Risk-Adaptive Learning (HOERAL) uses Bayesian optimization to adaptively weight outputs from different classifiers for improved generalization and confidence. Last but not least, Real-Time Federated Learning with Cross-Camera Collaboration (RTFL-CCC) should enable privacy-compliant tracking across cameras and collaborative inference of threats. This integrated architecture has shown very good accuracy (up to 95.8%) to detecting and localizing threats while gaining substantially on action co-occurrence, relational reasoning, and false-positive mitigation toward the process. The proposed system thus sets a strong, scalable, and ethically compliant paradigm for real-time anomaly detection in public safety surveillance settings.