Enhanced Intelligent Video Monitoring using Hybrid Integration of Spatiotemporal Autoencoders and Convolutional LSTMs
Ankita Umale-Nagmote, Charu Goel, Nidhi Lal · Informatica · 2025
Perceiving meaningful activities in surveillance videos presents significant challenges due to the ambiguous nature of anomalies and scene complexity. This paper proposes a hybrid deep learning framework that combines spatial-temporal autoencoders with convolutional LSTMs for automated anomaly detection in surveillance videos. The architecture integrates stacked convolutional autoencoders for semi-supervised feature representation with LSTM networks for preserving temporal information. Experiments conducted on the UCSD Ped1 dataset demonstrate that our LSTM-based Stacked CAE achieves an AUC of 83.5%, a detection rate of 81.5%, and an Equal Error Rate (EER) of 19.2%. The model particularly excels in temporal pattern recognition with an accuracy of 84.5% and sequence processing efficiency of 82.7%. Comparative analysis with state-of-the-art methods reveals that the proposed architecture achieves competitive performance, particularly in handling complex motion patterns and maintaining temporal consistency. The model shows significant improvement in false alarm rate reduction at 15.8% compared to the basic CAE’s 17.2%. The results demonstrate that integrating LSTM with stacked convolutional autoencoders provides a robust framework for real-world surveillance applications, especially in scenarios requiring both spatial and temporal anomaly detection.