Automatic Railway Traffic Object Detection System Using Feature Fusion Refine Neural Network under Shunting Mode

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Engineering

Background:

• Train accidents frequently occur during shunting mode at speeds below 45 km/h.
• Current manual observation methods by train attendants are prone to human error and delays.

Purpose of the Study:

• To develop an automated object detection system for enhanced safety in train shunting mode.
• To improve the real-time detection of railway conditions ahead of the train.

Main Methods:

• Proposed a Feature Fusion Refine neural network (FR-Net) utilizing convolutional neural networks (CNNs).
• Implemented a system with depthwise-pointwise convolutions, a coarse detection module, and an object detection module.
• Optimized anchor refinement and classification for accurate object localization and identification.

Main Results:

• FR-Net achieved a mean Average Precision (mAP) of 0.8953.
• The system demonstrated a real-time performance of 72.3 Frames Per Second (FPS) on a GeForce GTX1080Ti.
• The model achieved this performance with an input image size of 320x320 pixels.

Conclusions:

• FR-Net offers an effective balance between detection accuracy and real-time processing speed.
• The proposed system meets the practical requirements for automated object detection in train shunting operations.
• This technology has the potential to significantly reduce accidents in low-speed train operations.