Driving-Pattern Identification and Event Detection Based on an Unsupervised Learning Framework: Case of a Motorcycle-Riding Simulator

Analysis of human driving behavior aims to inspect drivers’ behavior in the real-world and in a virtual environment. The study of driving behaviors can be conducted in naturalistic situations or controlled experiments. Analyzing driving behaviors based on the data collected in naturalistic driving experiments or controlled experiments in the real-world or in a virtual environment is beneficial to fill in many of the knowledge gaps about driving behaviors and risk factors. The amount of data collected during complex experiments with many laps and many drivers tested under different experimental conditions and with different instructions can be huge. Analyzing such data can thus be considered challenging and time-consuming if done manually because it requires calling on experts in traffic psychology to inspect and understand various specific situations at a macroscopic scale involving different riders and at a microscopic scale for a particular rider on a specific lap. Also, it can be challenging in an unsupervised context to detect and match the same patterns in different laps to study similar patterns and spot important and risky events. This paper proposes a multi-step framework for analyzing driving behavior on both the macroscopic and microscopic scales. The core step of this framework is based on unsupervised machine learning algorithms applied to driving-pattern identification and the detection of critical driving events using anomaly-detection algorithms. The detected events are interpreted and described by computing their feature importance using graphs centrality measures. This provides new insight into driving behavior by identifying the motives behind the driver’s actions. The present experimental study, based on a dataset collected from the Honda Riding Trainer (HRT) simulator was conducted in the context of the European project SimuSafe and demonstrates the effectiveness of the proposed methodology. These results argue in favor of the development of such methodologies in driving-behavior studies.