DeepGuard: Enhancing Violence Detection in Smart Cities Through Deep Learning

Abstract

DeepGuard is a deep-learning-based violence detection system designed to enhance public safety in smart-city surveillance environments. The system uses computer vision and a hybrid neural architecture—a Convolutional Neural Network (ResNet18) for spatial feature extraction combined with a Bi-directional Long Short-Term Memory (Bi-LSTM) network for temporal modelling—to analyse video footage from CCTV and uploaded sources. Frames are sampled at fixed intervals, resized, normalised, and converted to tensors before being passed through the hybrid model, which classifies the sequence as Violence or Non-Violence. When a violent activity is detected, the system triggers an immediate alert to designated personnel through an interactive Streamlit-based interface. The implementation is realised in Python using PyTorch, OpenCV, NumPy, and Pillow. By combining spatial and temporal cues, the system reduces dependence on continuous manual monitoring and provides a scalable, automated solution suitable for integration with existing smart-city surveillance infrastructure. Functional, integration, system, performance, and accuracy tests confirm the correctness of all modules and end-to-end behaviour of the prototype.