Multi-Channel Generative Framework and Supervised Learning for Anomaly Detection in Surveillance Videos.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Current anomaly detection methods often rely on reconstruction errors from normal samples.
• Existing approaches using apparent motion and appearance reconstruction have limitations in distinguishing complex anomalies.
• The need for improved feature representation and the potential of supervised learning in anomaly detection are recognized.

Purpose of the Study:

• To propose a flexible multi-channel framework for generating diverse frame-level features.
• To investigate the efficacy of supervised learning for enhancing anomaly detection performance.
• To achieve state-of-the-art results in frame-level anomaly detection and localization.

Main Methods:

• Developed a multi-channel framework utilizing four Conditional Generative Adversarial Networks (CGANs) for feature generation.
• Employed Peak Signal-to-Noise Ratio (PSNR) to encode differences between generated and ground-truth information.
• Applied supervised learning with a Support Vector Machine (SVM) classifier on generated features, incorporating abnormal samples for training.
• Utilized Mask R-CNN for object-centric anomaly localization.

Main Results:

• The proposed PSNR features combined with supervised SVM outperformed previous error map methods.
• Achieved state-of-the-art frame-level Area Under the Curve (AUC) on the Ped1 and ShanghaiTech datasets.
• Demonstrated a significant performance improvement of up to 9% over unsupervised strategies on the ShanghaiTech dataset.

Conclusions:

• The multi-channel framework effectively generates discriminative features for anomaly detection.
• Supervised learning significantly boosts anomaly detection performance compared to unsupervised methods.
• The proposed approach represents a substantial advancement in video anomaly detection and localization.