A deep learning based detection algorithm for anomalous behavior and anomalous item on buses.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Safety

Background:

• Public transportation safety is a growing concern, necessitating advanced surveillance and anomaly detection methods.
• Existing systems often struggle with real-time analysis of complex passenger behaviors and diverse abnormal objects within confined spaces like buses.

Purpose of the Study:

• To develop and validate a novel algorithm for detecting abnormal passenger behavior and objects on buses.
• To enhance the accuracy and timeliness of anomaly detection in bus surveillance systems.
• To create a practical, embedded system for real-world application on buses.

Main Methods:

• Established a comprehensive library of abnormal passenger behaviors and objects specific to bus environments.
• Developed a mask detection and abnormal object detection and analysis (MD-AODA) algorithm, enhancing the YOLOv5 deep learning model.
• Integrated onboard face detection with target tracking for accurate face mask detection.
• Employed a geometric scale conversion approach for effective detection of large abnormal objects.
• Designed an embedded video analysis system to deploy the algorithm on actual bus data.

Main Results:

• The proposed MD-AODA algorithm demonstrated improved accuracy and timeliness in detecting anomalies compared to existing methods.
• Experiments using real bus video data validated the algorithm's effectiveness and practical applicability.
• The embedded system successfully integrated the algorithm for real-time analysis.

Conclusions:

• The developed MD-AODA algorithm and embedded system offer a practical and effective solution for enhancing bus safety.
• The strategy provides a significant advancement in automated anomaly detection for public transportation.
• The findings confirm the algorithm's validity and potential for widespread adoption in intelligent transportation systems.