Enhanced ConvLSTM with Hierarchical Attention Scheme for Unusual Activity Detection in Classrooms or Laboratories

Surveillance camera systems have been essential in enhancing security protocols, serving as crucial tools for monitoring and deterring unauthorized or unusual activities across various settings. This paper addresses the significant vulnerability of educational spaces to numerous security threats, such as unauthorized access, theft, vandalism, and risks to the safety of students and staff. In an age where both physical and digital security are critically important, there is an increasing demand for proactive surveillance solutions that can quickly and accurately identify anomalies and potential security breaches. Deep learning techniques have demonstrated remarkable effectiveness in identifying patterns, anomalies, and subtle deviations in behavior within real-time video feeds. We have developed an enhanced ConvLSTM model incorporating spatial and temporal attention mechanisms to extract essential features from raw video footage, which are then processed by the backbone network. The architecture is designed to selectively focus on the most valuable features, thereby reducing computational complexity and enhancing the model's detection performance. The model is fine-tuned to detect and classify activities into usual and unusual categories. The proposed design is trained on videos from the UCF-Crime dataset and fine-tuned on a custom dataset created from YouTube videos, which includes videos of actions that are unusual or anomalous in educational environments. Our approach achieved approximately 99% accuracy on the UCF-Crime dataset, surpassing detection baselines (85.53% to 88.56%) and outperforming other architectures such as C3D, ConvLSTM, and CNN+RNN models with and without attention mechanisms. Evaluation on a custom YouTube dataset yielded a detection accuracy of 82.66%.