A railway intrusion detection method based on decomposition and semi-supervised learning for accident protection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Safety

Background:

• Video surveillance is crucial for railway intrusion detection.
• Challenges include background similarity, lack of diverse training data, overfitting, small/distant target detection, and extreme data imbalance.
• Existing methods struggle with generalization and accurate detection in complex railway environments.

Purpose of the Study:

• To develop an effective and accurate railway intrusion detection method.
• To address challenges of background interference, data imbalance, and limited labeled data.
• To improve the generalization capacity of detection models.

Main Methods:

• Combining low-rank and sparse decomposition (LRSD) to separate background and foreground.
• Utilizing semantic segmentation to mask track regions in the background.
• Applying Semi-supervised Support Vector Domain Description (Semi-SVDD) using both labeled and unlabeled foreground data.

Main Results:

• The proposed method effectively removes background interference.
• Integration of labeled and unlabeled data significantly enhances detection performance.
• Numerical results demonstrate superior performance compared to benchmark methods.

Conclusions:

• The combined LRSD and Semi-SVDD approach offers a robust solution for railway intrusion detection.
• Effectively handling background clutter and data imbalance leads to improved accuracy and generalization.
• This method presents a promising advancement in railway security surveillance.