Hybrid CNN-LSTM Model for Jail Activity Recognition

Jail Activity Recognition (JAR), a task in the field of security and surveillance requires a comprehensive exploration strategy. Detecting critical activities could help officials make initial attempts to effectively manage jail environments. The study focuses on recognizing three targeted activities activities: fighting, shooting, and punching that might lead to more unfortunate escalated incidents. The paper proposes a hybrid approach to JAR that combines Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM). Model is trained on targeted videos taken from public domain. By leveraging the strengths of deep learning, this study attempts to capture spatial features and model temporal dependencies to detect critical activities in a crowded setting. The comparison results show that the proposed hybrid CNN-LSTM model has better performance in accuracy and robustness. Model detects subtle variations and is able to effectively handle real-world surveillance data challenges. The code for our proposed model is publicly available at GitHub link, to ensure transparency and reproducibility for researchers and practitioners in the field.