Authors: Sakshi Keshri, Nitin Namdev
Abstract: This paper presents a comprehensive review of deep learning techniques designed to enhance violence detection in surveillance systems. With the rapid advancement of surveillance technologies, the accurate identification of violent activities has become crucial for ensuring public safety. Conventional approaches often fail to cope with the complexity of video data, which inherently involves both spatial and temporal dynamics. To overcome these limitations, modern deep learning models such as Convolutional Neural Networks (CNNs), InceptionV3, Long Short-Term Memory (LSTM) networks, and hybrid architectures have been widely adopted. These methods excel at capturing spatial representations while simultaneously modeling temporal dependencies, making them well-suited for real-time violence detection tasks. The review further discusses essential preprocessing strategies—including noise reduction, feature extraction, and data augmentation—that significantly improve model robustness. In addition, it outlines persistent challenges such as class imbalance, scalability issues, and high computational costs, which remain key barriers to practical deployment
