Spatio-temporal prediction and reconstruction network for video anomaly detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing anomaly detection models often struggle with either generating significant errors for anomalies (reconstruction-based) or being sensitive to noise (future frame prediction).
• Many current methods utilize single-scale information, limiting spatial feature extraction and neglecting temporal continuity, thus impacting anomaly detection accuracy.
• Objects in videos appear at various scales, necessitating methods that can capture features across different sizes.

Purpose of the Study:

• To propose a novel method for improved video anomaly detection.
• To address the limitations of single-scale feature extraction and lack of temporal continuity in existing methods.
• To enhance the accuracy of anomaly detection in complex video scenarios.

Main Methods:

• Utilizing a hybrid dilated convolution (HDC) module to extract comprehensive spatial features by employing different receptive fields for objects of various scales.
• Implementing a deeper bidirectional convolutional long short-term memory (DB-ConvLSTM) module to effectively capture and remember temporal information between consecutive video frames.
• Integrating multi-scale spatial feature extraction with robust temporal modeling for anomaly detection.

Main Results:

• The proposed method demonstrates superior performance in detecting abnormalities compared to state-of-the-art methods.
• Experiments show improved anomaly detection accuracy across various video scenes.
• The combined approach effectively balances the strengths of reconstruction and prediction models while mitigating their weaknesses.

Conclusions:

• The novel method, leveraging HDC and DB-ConvLSTM, significantly enhances video anomaly detection performance.
• Capturing multi-scale spatial features and temporal continuity is crucial for accurate anomaly detection.
• This approach offers a more robust solution for identifying abnormalities in complex video data.