Deep BiLSTM Attention Model for Spatial and Temporal Anomaly Detection in Video Surveillance.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Manual monitoring of increasing surveillance cameras is challenging.
• Automated anomaly detection systems are crucial for public safety.
• Need for real-time identification of diverse abnormal events (accidents, fires, etc.).

Purpose of the Study:

• To propose the Composite Recurrent Bi-Attention (CRBA) model for enhanced anomaly detection in surveillance videos.
• To improve the accuracy and efficiency of identifying abnormal events in real-time.
• To address both spatial and temporal challenges in video surveillance analysis.

Main Methods:

• Utilized DenseNet201 for robust spatial feature extraction.
• Employed BiLSTM networks to capture temporal dependencies across video frames.
• Incorporated a multi-attention mechanism to focus on critical spatiotemporal regions.

Main Results:

• The CRBA model demonstrated high accuracy in detecting anomalies.
• Achieved strong performance on the University of Central Florida (UCF) and Road Anomaly Dataset (RAD).
• Effectively distinguished between normal and abnormal behaviors in surveillance footage.

Conclusions:

• The CRBA model offers improved detection and classification of anomalies.
• Enhances resource efficiency and minimizes response times in critical security situations.
• Provides an invaluable tool for public safety and security operations requiring rapid, accurate responses.