Development of a data driven method for determining road conditions using vehicle-mounted sensors

Abstract

The maintenance of the road surface is one of the most impactful variables to ensure the safety of all tra!c participants. Although many new assistantsystems have been developed, they all rely on the contact between the tires and the road surface to work. In this thesis a system is developed that uses costfriendly IMU data, transformed to the time-frequency domain, as input for an unsupervised one dimensional convolutional auto-encoder. The latent space is then used as input for the k-means clustering. Camera images recorded by one of the participating vehicles and satellite images are then used to evaluate the performance of the system. The result are corridors of varying length which are sorted into one of four clusters each representing a di↵erent road segment condition.