Small-Scale Urban Object Anomaly Detection Using Convolutional Neural Networks with Probability Estimation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Urban Security

Background:

• Automated analysis of urban surveillance data is crucial due to the increasing number of cameras.
• Identifying anomalous events in real-time is essential for effective security and public safety.

Purpose of the Study:

• To develop and validate a methodology for detecting anomalous events in urban surveillance sequences.
• To leverage pre-trained Convolutional Neural Networks (CNN) and Super-Resolution (SR) for anomaly detection.

Main Methods:

• An offline stage using a pre-trained CNN to identify common element locations and create a density matrix.
• An online stage where the CNN uses the density matrix to assess real-time anomaly probabilities.
• Utilizing super-resolution models to enhance image quality for analysis.

Main Results:

• The methodology effectively detects various anomalies, including unusual pedestrian routes.
• Experimental results confirm the approach's efficacy in real-world urban surveillance scenarios.
• The density matrix accurately captures spatial patterns of common elements.

Conclusions:

• The proposed method offers a practical and reliable solution for anomaly detection in urban environments.
• This contributes to improved public safety and proactive urban management.
• Timely anomaly detection enables faster reactions and enhances overall urban safety.