Railway Intrusion Risk Quantification with Track Semantic Segmentation and Spatiotemporal Features

Area of Science:

• Artificial Intelligence
• Railway Engineering
• Computer Vision

Background:

• Foreign object intrusion in railway areas presents significant safety risks.
• Current visual detection methods lack quantitative risk assessment capabilities.

Purpose of the Study:

• To develop a railway intrusion risk quantification method integrating track semantic segmentation and spatiotemporal features.
• To enhance train operation safety through quantitative risk assessment and graded early warnings.

Main Methods:

• Utilized an improved BiSeNetV2 network for accurate track region extraction.
• Constructed physical-constrained risk zones based on railway structure gauge standards.
• Developed a lightweight detection architecture with a Dilated Transformer module for improved accuracy, especially for small objects.
• Integrated object category weights, lateral risk coefficients, longitudinal distance decay, and velocity compensation for comprehensive risk assessment.

Main Results:

• Achieved 84.9% mean average precision (mAP) on a proprietary dataset.
• Outperformed baseline models by 3.3% in intrusion detection accuracy.
• Demonstrated the capability for quantitative intrusion risk assessment and graded early warning.

Conclusions:

• The proposed method enables data-driven decision support for active train protection systems.
• Significantly enhances intelligent railway safety protection capabilities by combining lateral distance detection with multidimensional risk indicators.