An Intelligent Obstacle Detection Method for Rail Transit Scenarios

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Engineering

Background:

• Traditional railway monitoring systems lack real-time foreign object detection capabilities.
• Ensuring train operational safety requires advanced obstacle identification methods.

Purpose of the Study:

• To develop an intelligent visual sensing approach for detecting foreign objects in railway track regions.
• To enhance the accuracy and efficiency of obstacle detection using an improved YOLOv8 model.

Main Methods:

• The ACX-YOLOv8 model integrates spatial coordinate attention mechanism (SCAM) and cascaded dilated convolution (CDConv).
• An X6 decoupled detection head is employed to improve detection accuracy and inference speed.
• The model is built upon the YOLOv8 framework for object detection.

Main Results:

• The ACX-YOLOv8 model achieved a mean average precision (mAP50) of 87.1%, a 2.7% improvement over the baseline YOLOv8.
• The model demonstrated a 1.8% mAP50 increase on the PASCAL VOC dataset, verifying its generalization ability.
• The optimized model has 4.85 million parameters, balancing lightweight properties with high detection precision.

Conclusions:

• ACX-YOLOv8 provides rapid and accurate detection of foreign objects in railway tracks.
• The algorithm offers reliable technical support for real-time obstacle monitoring in complex railway environments.
• This contributes to enhancing the safety and intelligentization of railway operations.