Statistical Hypothesis Detector for Abnormal Event Detection in Crowded Scenes.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Abnormal event detection in crowded scenes is challenging.
• Existing methods struggle with divergent normal patterns and underutilize abnormal event patterns.
• Lack of labeled abnormal data hinders training.

Purpose of the Study:

• To propose a new abnormality detector for crowded scenes.
• To address limitations of existing methods by utilizing abnormal event patterns and handling pattern divergence.
• To improve the accuracy and robustness of abnormal event detection.

Main Methods:

• Employing a statistical hypothesis test for abnormality detection.
• Treating each sample as a combination of event patterns.
• Adaptively extracting abnormal patterns from unlabeled testing samples.
• Approximating complex noise distributions using a mixture of Gaussians.
• Implementing an online updating strategy for normal event patterns.

Main Results:

• The proposed detector identifies abnormal events by detecting abnormal patterns and high abnormality scores.
• Using a mixture of Gaussians improves video event modeling and detection accuracy.
• The online updating strategy effectively handles divergent normal events, reducing false detections.
• Extensive experiments validate the algorithm's effectiveness against state-of-the-art methods.

Conclusions:

• The proposed statistical hypothesis test-based abnormality detector is effective for crowded scenes.
• The method successfully addresses challenges related to pattern divergence and data scarcity.
• The approach demonstrates improved accuracy and reduced false positives in abnormal event detection.