Dual Weakly Supervised Anomaly Detection and Unsupervised Segmentation for Real-Time Railway Perimeter Intrusion...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Railway Engineering

Background:

• High-speed train operations necessitate advanced track intrusion detection systems.
• Onboard systems face limitations due to computational constraints and motion blur.
• Effective monitoring of track perimeters is crucial for safety.

Purpose of the Study:

• To develop an intelligent monitoring system for foreign objects on high-speed train tracks.
• To address challenges in dataset annotation and unidentified target detection.
• To enhance real-time anomaly detection and localization capabilities.

Main Methods:

• Implemented weakly supervised video anomaly detection and unsupervised foreground segmentation.
• Utilized Xception3D pretraining and attention mechanisms for feature extraction.
• Employed Top-K sample selection, amplitude score/feature loss, and time-smoothing for anomaly discrimination.
• Applied a multiscale variational autoencoder for foreign object localization.
• Engineered a pixel-level background weight distribution loss function.

Main Results:

• Video anomaly detection model achieved an Area Under the Curve (AUC) of 0.99.
• The system processes 2-second video segments in 0.41 seconds.
• Foreground segmentation algorithm achieved an F1 score of 0.9030 on the track anomaly dataset and 0.8375 on CDnet2014.
• Achieved a processing speed of 91 Frames Per Second (FPS) for foreground segmentation.

Conclusions:

• The developed system effectively monitors foreign objects on high-speed train tracks.
• Weakly supervised learning and advanced deep learning techniques improve detection accuracy and efficiency.
• The system demonstrates high performance and real-time processing capabilities, confirming its practical utility.