Pavement Transverse Cracking Detection Based on the Vehicle’s Vertical Acceleration Performed Via Transfer Learning and Wavelet Scattering Transform

Road surface anomalies can damage pedestrians, vehicles, and vehicle users. The above has motivated using signal processing and machine learning algorithms to develop systems that automatically detect road surface anomalies by processing the vehicle’s vibrations while it travels over the road surface. However, machine learning algorithms require a sizable sample size to train them, which complicates generating road surface anomaly detection systems. This paper uses a small sample size to compare the performance of Transfer Learning (TL) and the Wavelet Scattering Transform (WST) for pavement transverse cracking detection based on vehicle vertical acceleration data. First, TL was applied by fine-tuning the pre-trained SqueezeNet with scalograms of vehicle vertical acceleration signals generated via the continuous wavelet transform and Generalized Morse Wavelets. On the other hand, the WST coefficients obtained from the accelerations signals were used to train a regularized linear discriminant analysis (LDA). The results of this study showed that the fine-tuned SqueezeNet provided better performance than the regularized LDA trained with WST features in terms of accuracy, sensitivity, specificity, and F1-score at the cost of a higher training time and a more complex classifier.