Discerning truck stop semantics through latent space clustering

Sammanfattning: GPS systems have been in use for navigational purposes for almost three decades and have found their way into location tracking systems. The principal, Scania has a fleet of 300 thousand connected vehicles sending in position information. In this paper, we make use of position information sourced from Scania’s connected trucks, which have been abstracted into stops. The stop abstractions are built by coalescing raw position information on the basis of temporal and spatial thresholds and denote locations where the truck halts. We apply unsupervised machine learning approaches to try and understand the semantics behind these stops. The features of the truck stops are projected into a low dimensional latent space using deep autoencoders, and a clustering objective is then optimized in this low dimension space. The resultant clusters are found to be representative of different types of truck stops. The characterized truck stoppages can be useful for understanding the truck usage patterns as well transport hub usage statistics.