Video Anomaly Detection Based on Convolutional Recurrent AutoEncoder.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video anomaly detection is crucial for surveillance and scene understanding but faces challenges from complex backgrounds and scene changes.
• Existing methods struggle with accurately defining and identifying anomalies in dynamic video data.
• The core idea is that normal video patterns have lower reconstruction errors than anomalous ones.

Purpose of the Study:

• To develop an advanced method for video anomaly detection.
• To address the challenges posed by complex scenes and temporal irregularities in videos.
• To improve the accuracy and reliability of anomaly identification in computer vision tasks.

Main Methods:

• Proposed a Convolutional Recurrent AutoEncoder (CR-AE) integrating an attention-based Convolutional Long-Short-Term Memory (ConvLSTM) and a Convolutional AutoEncoder.
• Utilized ConvLSTM to capture temporal pattern irregularities and Convolutional AutoEncoder for spatial irregularities.
• Employed an attention mechanism to enhance feature extraction from hidden states and a convolutional decoder for reconstruction.

Main Results:

• The CR-AE model demonstrated strong performance on the UCSD ped2 and Avenue datasets.
• Achieved a frame-level Area Under the Curve (AUC) of 95.6% on the UCSD ped2 dataset.
• Obtained a frame-level AUC of 73.1% on the Avenue dataset, showcasing effectiveness in diverse scenarios.

Conclusions:

• The proposed CR-AE method is effective for video anomaly detection.
• The combination of temporal and spatial feature extraction with attention mechanisms improves anomaly identification.
• CR-AE offers a robust solution for real-world applications like video surveillance and traffic analysis.