A Background-Agnostic Framework With Adversarial Training for Abnormal Event Detection in Video.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Abnormal event detection in videos is challenging due to the context-dependent nature of rare events.
• Existing methods often struggle with variations across different scenes and backgrounds.

Purpose of the Study:

• To propose a background-agnostic framework for abnormal event detection.
• To develop a method that learns solely from normal events, overcoming the scarcity of abnormal event data.
• To enhance the robustness and applicability of abnormal event detection systems across diverse scenarios.

Main Methods:

• A framework integrating object detection, appearance and motion auto-encoders, and classifiers.
• An adversarial learning strategy using out-of-domain pseudo-abnormal examples for auto-encoder training.
• A segmentation branch to ensure focus on main objects within bounding boxes.

Main Results:

• The proposed framework demonstrates favorable performance compared to state-of-the-art methods on four benchmark datasets.
• Empirical results validate the effectiveness of the background-agnostic approach and adversarial learning strategy.
• The method shows strong generalization capabilities across different scenes.

Conclusions:

• The developed background-agnostic framework offers a robust solution for abnormal event detection.
• Adversarial learning effectively addresses the challenge of limited abnormal event data.
• The approach provides a significant advancement in video anomaly detection research.