Anomaly Detection in Video Surveillance Using a Novel BiLSTM-Hybrid Temporal Encoder
3rd International Conference on Artificial Intelligence, Computer, Data Sciences, and Applications, ACDSA 2026, Boracay Island, Philippines, 5 - 07 February 2026, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/acdsa67686.2026.11467910
- City: Boracay Island
- Country: Philippines
- Keywords: Anomaly detection, Deep learning, Smart city, Spatialtemporal modeling, Surveillance videos, Temporal encoder
- Middle East Technical University Affiliated: Yes
Abstract
Ensuring reliable anomaly detection in surveillance videos is vital for smart city safety, but conventional approaches often struggle to capture both long-term temporal dynamics and discriminative spatial patterns. This work introduces a novel BiLSTM-Hybrid temporal encoder (BHTE), which integrates bidirectional LSTM layers to model forward and backward temporal dependencies with either convolutional or dense layers for refining feature representations. The CNN-based variant emphasizes localized spatiotemporal feature extraction, whereas the DNN-based variant focuses on compact abstraction of temporal embeddings. By combining recurrent and feedforward components in this manner, the framework effectively balances temporal context modeling with feature refinement. We validate the proposed model on the Smart-City CCTV Violence Detection Dataset (SCVD). The experimental results show that the DNN variant achieves 93% test accuracy, while the CNN variant achieves a higher accuracy of 96.34%, outperforming the conventional sequence learning baselines. These findings highlight the robustness and scalability of the BHTE framework, making it a promising solution for real-world intelligent surveillance and video-based anomaly detection.