Detection Anomaly in Video Based on Deep Support Vector Data Description.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video surveillance is crucial for public safety, but detecting abnormal events remains challenging due to ambiguous definitions and data scarcity.
• Current deep learning methods often rely on generic network structures, not optimized for anomaly detection.

Purpose of the Study:

• To develop an effective and accurate method for video anomaly detection.
• To address the limitations of existing deep learning approaches in identifying unusual events in surveillance footage.

Main Methods:

• Proposed a Deep Support Vector Data Description (DSVDD) method.
• Utilized a deep neural network to map normal data samples into the smallest possible hypersphere.
• Classified samples inside the hypersphere as normal and outside as abnormal.

Main Results:

• Achieved 86.84% frame-level AUC on the CUHK Avenue dataset.
• Achieved 73.2% frame-level AUC on the ShanghaiTech Campus dataset.
• Demonstrated superior performance compared to existing state-of-the-art methods.

Conclusions:

• The proposed DSVDD method offers a robust solution for video anomaly detection.
• This approach effectively learns data representations and normal models for accurate abnormality identification.