Smartphone Sensor Data Augmentation for Automatic Road Surface Assessment Using a Small Training Dataset

Smartphones equipped with motion sensors are widely used for data collection in research aimed at the establishment of smart transportation and at, more specifically, automatic road condition assessment. To perform the assessment task, machine learning classifier systems are developed to analyze patterns of vibration signals recorded from a driver's smartphone. Obtaining a balanced training dataset required for the classifier system to work properly is, however, a difficult task. The presented study develops an approach based on an Unrolled Generative Adversarial Network (Unrolled GAN) to produce synthetic data for balancing the training dataset. Experiments conducted in the study demonstrated that the approach allows for generating high-quality synthetic data as long as the unrolled GAN are kept controlled to balance the discriminator and generator modules of the networks.