End-to-End Train Horn Detection for Railway Transit Safety

Area of Science:

• Artificial Intelligence
• Signal Processing
• Railway Engineering

Background:

• Train horns are crucial audible warning signals for railway safety, particularly at barrier-free crossings.
• Early detection of train arrival is essential for preventing accidents and ensuring smooth traffic operations.
• Existing methods for sound classification may not be optimal for the complex traffic soundscape.

Purpose of the Study:

• To develop a deep learning system for early train arrival detection by recognizing train horn sounds.
• To propose and evaluate a novel two-stream Convolutional Neural Network (CNN) model, THD-RawNet, for audio classification.
• To assess the model's performance and robustness in various traffic and noise conditions.

Main Methods:

• A custom dataset of train horn sounds, car horns, and traffic noises was created for training and testing.
• A novel two-stream end-to-end CNN model (THD-RawNet) was developed, combining sequential 1D-CNN and multi-resolution 1D-CNN branches.
• The model processes raw audio waveforms to extract features for train horn detection (THD).

Main Results:

• The THD-RawNet model demonstrated superior accuracy and robustness compared to baseline models.
• Achieved 95.11% accuracy on test data under normal traffic conditions with a one-second input.
• Maintained over 93% accuracy in noisy conditions with -10 dB Signal-to-Noise Ratio (SNR).

Conclusions:

• The proposed THD-RawNet model is effective for early train horn detection in complex soundscapes.
• The two-stream architecture and multi-resolution feature extraction enhance classification performance.
• The system can be integrated into smart railway crossings, private cars, and autonomous vehicles to improve transit safety.