REAL-TIME RAILWAY TRACK OBSTACLE DETECTION USING RESNET18 AND MULTI-HEAD ATTENTION ON RADAR SPECTROGRAMS

Abstract

Railway track safety is important to the correct identification of items on the tracks and the exact estimate of their distances. Traditional detection systems often suffer with real-time processing, noise interference, and low accuracy, which might jeopardise safety measures. To address these issues, this research offers a unique deep learning system built on a ResNet-18 backbone and a dual-head attention mechanism that allows for simultaneous object categorisation and distance estimates using radar data. Objects are recognised when transmitted radar waves reflect back after striking a target; in the absence of an object, the waves proceed without reflection. The recovered radar signals are then recorded and processed into spectrograms using the Short-Time Fourier Transform (STFT), which allows for accurate time-frequency representation of the signals.

These spectrograms feed into the proposed model, where the dual-head attention mechanism enables focused learning for both object presence categorisation and distance estimate via regression. The system was trained and evaluated on a realistically generated dataset that included ambient noise and item diversity to mimic real-world railway settings. The experimental findings showed excellent classification ability, with accuracy, precision, recall, and F1 scores ranging from 0.98 to 0.99. The regression head demonstrated strong distance estimation capabilities, with a Mean Squared Error (MSE) of 5320.30, Mean Absolute Error (MAE) of 25.19, and an R-squared value of 0.6041. This paper introduces the novel approach of using a dual-head attention framework on radar spectrogram data to address classification and regression tasks simultaneously, resulting in a robust, efficient, and accurate solution that significantly improves automated railway track monitoring and contributes to improved railway safety.