Smart Wearable-Based Behavioural Anomaly Detection Framework for Women's Safety Using AI

Abstract

Personal safety remains a pressing concern for women worldwide: the World Health Organization estimates that nearly one in three women globally (31.6%, roughly 840 million women and girls) have experienced physical and/or sexual violence in their lifetime [1]. Reactive panic-button wearables depend on the wearer being able to consciously trigger an alert, which is frequently impossible during an assault. This paper presents an AI-assisted wearable safety framework that continuously and passively monitors behavioural and physiological signals - tri-axial accelerometer, gyroscope, heart rate and galvanic skin response (GSR) - to detect deviations from a person's normal activity pattern, and fuses this with a YOLOv8n-based environmental object-recognition module operating on associated camera frames. A compact feed-forward neural network ("WearNet": 33 input features, 4,289 trainable parameters) is trained to separate four normal activities (sitting, standing, walking, running) from four distress/anomalous events (falls, physical struggle, panic flight, post-impact immobility) on a physiologically-informed synthetic benchmark of 1,336 four-second windows drawn from 46 simulated subjects. On a subject-disjoint held-out test set (10 unseen subjects, 281 windows), WearNet attains 99.64% accuracy, ROC-AUC = 0.998, PR-AUC = 0.998 and a single misclassification (zero false positives, one false negative). The YOLOv8n module, evaluated with a genuine ONNX/onnxruntime inference pipeline on real photographs, detects people and objects at a mean CPU latency of 82.4 ms (~12.1 FPS) from a 12.18 MB model. We report calibration, decision-curve and permutation feature-importance analyses, benchmark against five alternative classifiers including an unsupervised, normal-only Isolation Forest, and - importantly - discuss why controlled/synthetic benchmarks such as this one are expected to over-state real-world performance, drawing on published wearable fall-detection literature for context.