CNN and Bilstm Based Framework for Real Life Violence Detection from CCTV Videos
Shrey Aggarwal, Rishabh Ranjan, Manasvi Sinha, Vipin Pal, Riti Kushwaha · 2024
This paper focuses on developing an automated violence detection system for CCTV footage using machine learning techniques. The dataset comprises real-life videos cap-turing instances of violence and non-violent activities in various public spaces. Extensive preprocessing techniques are applied to enhance the quality and diversity of the dataset, including frame extraction, augmentation, and image enhancement. The pro-posed solution adopts a hybrid architecture, integrating a pre- trained MobileNetV2 model with a Bidirectional Long Short- Term Memory (BiLSTM) layer for spatial and temporal feature extraction. Model compilation and training are conducted with appropriate optimization algorithms and callbacks, resulting in a high-performing violence detection system. Performance evaluation metrics demonstrate the efficacy of the developed model, achieving an accuracy rate of 96 % on the test dataset. Comparison with existing techniques underscores the compet-itiveness of the proposed solution.