Local distinguishability aggrandizing network for human anomaly detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Surveillance systems increasingly require intelligent methods for detecting abnormal events.
• Existing anomaly detection methods suffer from network distraction and limited discriminating ability.

Purpose of the Study:

• To propose a novel supervised network, LDA-Net, to address shortcomings in current video anomaly detection.
• To enhance the network's ability to distinguish between normal and abnormal human behaviors.

Main Methods:

• Implemented a Local Distinguishability Aggrandizing Network (LDA-Net) with human detection and anomaly detection modules.
• Focused on segmented human subject patches to learn individual motion characteristics.
• Utilized a dual sub-branch anomaly detection module for joint anomaly detection and action recognition.
• Introduced a novel inhibition loss function to mitigate misclassification in imbalanced datasets.

Main Results:

• LDA-Net achieved state-of-the-art performance on public benchmark datasets.
• Demonstrated effectiveness in both frame-level and pixel-level anomaly detection tasks.
• Showcased superior results on the UCSD Ped2 and Subway Exit datasets.

Conclusions:

• The proposed LDA-Net effectively detects and locates anomalous behaviors in surveillance videos.
• Joint anomaly detection and action recognition enhance feature extraction for distinguishing behaviors.
• The novel inhibition loss function improves performance on imbalanced datasets, leading to state-of-the-art results.