A Fast Intrusion Detection Method for High-Speed Railway Clearance Based on Low-Cost Embedded GPUs

Area of Science:

• Computer Science
• Artificial Intelligence
• Railway Engineering

Background:

• Current railway intrusion detection systems are often costly or have limited coverage, hindering large-scale application.
• High computational demands of existing deep learning methods limit their real-time effectiveness.

Purpose of the Study:

• To develop a fast, low-cost, and effective railway intrusion detection solution for high-speed railways.
• To address the computational burden and improve the detection accuracy of small objects in railway environments.

Main Methods:

• A novel neural network based on the Single Shot Detector (SSD) framework was developed.
• An improved MobileNet was utilized as the feature extractor, coupled with an efficient feature fusion module.
• Feature map weights were introduced via convolution to fuse multi-scale features, enhancing small object detection.

Main Results:

• The proposed method achieved 89% mean Average Precision (mAP) on a dedicated railway intrusion detection dataset.
• The system demonstrated an average processing time of 38.6 ms per frame on an NVIDIA Jetson TX2 GPU.
• TensorRT optimization enabled efficient deployment on a low-cost embedded platform.

Conclusions:

• The developed system offers a computationally efficient and accurate solution for real-time railway intrusion detection.
• The lightweight network architecture and feature fusion techniques contribute to improved performance.
• This approach provides a viable, cost-effective alternative for enhancing railway safety and security.