An Adaptive Track Segmentation Algorithm for a Railway Intrusion Detection System.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Engineering

Background:

• Video surveillance is crucial for railway intrusion detection.
• Manual track area labeling is inefficient and challenging, especially for PTZ cameras.
• Automated track segmentation requires algorithms with low memory and fast inference.

Purpose of the Study:

• To develop an adaptive segmentation algorithm for automatic track area delineation.
• To reduce computational burden and improve efficiency in video surveillance systems.
• To achieve accurate track segmentation with minimal hardware resources.

Main Methods:

• Image segmentation into fragmented regions using adaptive Gaussian kernels and Hough transformation.
• Clustering of fragmented regions into local areas based on boundary weight and size.
• Classification of local areas using a simplified Convolutional Neural Network (CNN) with pre-trained kernels and a diversity-enhancing loss function.

Main Results:

• The proposed algorithm effectively delineates track area boundaries with a light computational load.
• Experimental results demonstrate a balance between segmentation precision, calculation time, and hardware cost.
• The simplified CNN achieves fast and accurate classification of track areas.

Conclusions:

• The adaptive segmentation algorithm offers an efficient solution for automated track area labeling in railway surveillance.
• The method addresses the limitations of manual labeling and the demands of real-time detection.
• This approach contributes to more effective and cost-efficient railway intrusion detection systems.