Probabilistic memory auto-encoding network for abnormal behavior detection in surveillance video.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Abnormal behavior detection is crucial for intelligent surveillance systems, aiding in anti-terrorism and public security.
• A significant challenge is the extreme imbalance between normal and abnormal behavior data in datasets.
• Existing methods struggle with accurately identifying anomalies due to this data imbalance.

Purpose of the Study:

• To design a novel probabilistic memory model-based network for robust abnormal behavior detection.
• To address the data imbalance issue by learning the distribution of normal behaviors.
• To enhance the accuracy and reliability of intelligent surveillance systems.

Main Methods:

• Utilized an auto-encoding model as the backbone for feature extraction.
• Employed an autoregressive conditional probability estimation model and a normal distribution memory model as auxiliary modules.
• Incorporated causal 3D convolution and time-dimension shared fully connected layers to prevent future information leakage.

Main Results:

• The proposed network effectively learns normal behavior distributions, achieving accurate anomaly detection.
• The model demonstrated significant advantages in detecting abnormal behaviors compared to existing methods on public datasets.
• Ablation and comparison experiments validated the algorithm's effectiveness.

Conclusions:

• The probabilistic memory model-based network offers a promising solution for abnormal behavior detection in surveillance.
• The approach effectively handles data imbalance, leading to improved detection performance.
• This work contributes to advancing intelligent surveillance for enhanced security and safety.