Abstract:
Vehicle telemetry data is becoming more ubiquitous with increasingly sensorised vehicles, but making sense of the vehicles' purpose remains challenging without additional context. Clustering the vehicle activity data and identifying the underlying facilities where the activities occur reveals much insight, particularly for logistics planning. Unfortunately, current research typically only looks at a single point in time. This paper contributes by matching geospatial patterns, each representing a facility where trucks perform activities over multiple periods. The contribution is a necessary first step in studying how urban freight movement and its underlying inter-firm networks of connectivity change over time. We demonstrate how to overcome three challenges. Firstly, the complexity of identifying facilities from non-regular geometric polygons. Secondly, the challenge associated with the scale of comparing more than 200,000 facilities on a month-to-month basis over a multi-year period. Finally, overcoming the computational challenge of the workflow and getting the required performance on a consumer-grade laptop. The paper evaluates various machine learning algorithms, highlighting a SVM that outperforms more popular deep learning and neural network alternatives, with a mean average accuracy of 96.9 %.
