Revealing Trip Purposes in Raw GPS Data by Applying a Multi-Phase Clustering Approach to Semantic Trajectories

The availability of GPS data from motorized vehicles potentially enables a much more detailed, accurate, and up-to-date analysis of traffic behavior than survey data. Exact routes and driving behavior (speed, acceleration, etc.) can be observed, making it possible to analyze and model congestion and emissions on specific road sections. However, these advantages come with drawbacks: Raw GPS data usually only includes timestamps and coordinates and lacks information on trip purpose, transportation mode, and socio-demographics. This is one reason traffic simulation models are still primarily based on conventional travel survey data. Conventional travel surveys suffer from small sample sizes and infrequent updates. To exploit the potential of GPS data, we present a method for detecting trip purposes from raw GPS data of motorized vehicles. In contrast to previous approaches that rely on supervised machine learning and require ground truth data for training, our methodology uses only unsupervised machine learning techniques. This enables the utilization of any GPS datasets, a capability previously unavailable. The proposed method is based on the plausible assumption that especially destination areas can reveal trip purposes. We apply this methodology to analyze 170, 000 trips for Frankfurt / Germany. Initially, we segment the study area into clusters of sub-areas to condense origin and destination information, which we then integrate with other trip-related variables in a multi-phase trajectory clustering process. Our analysis identifies eight distinct trip purposes. These findings are validated by comparing them with results from the national travel survey, demonstrating the plausibility of our proposed methodology.