An Intelligent Deep Learning Framework for Violence Detection and Criminal Activity Identification in Smart Surveillance Systems

Authors: Shivam Namdev, Bhanu Pratap Singh

Abstract: The increasing growth of urbanization, public gatherings, and security threats has created a strong demand for intelligent surveillance systems capable of automatically detecting violent activities and identifying criminal behavior in real time. Traditional surveillance systems mainly rely on manual monitoring, which often leads to delayed response, human errors, and inefficient threat analysis. To address these limitations, this research proposes an Intelligent Deep Learning Framework for Violence Detection and Criminal Activity Identification in Smart Surveillance Systems using advanced Deep Learning architectures and video analytics techniques. The proposed framework integrates preprocessing, segmentation, feature extraction, and hybrid Deep Learning models including ResNet50, MobileNetV2, LSTM, and the proposed InceptionV3 + LSTM architecture for violence classification. The system processes surveillance videos by extracting spatial and temporal features to accurately distinguish violent and non-violent activities. The dataset consists of two categories, namely Violence and Non-Violence, containing real-world surveillance scenarios such as sports, crowd movement, eating, singing, and violent human interactions. Experimental analysis demonstrates that the proposed InceptionV3 + LSTM model achieved superior performance with improved training accuracy, validation accuracy, and reduced loss values compared with existing models. The framework effectively captures complex motion patterns, human interactions, and abnormal behaviors in surveillance videos while reducing false detections. The proposed intelligent surveillance system can be applied in smart cities, public transportation, educational institutions, stadiums, and crowded environments to improve public safety, automated threat detection, and real-time security monitoring.