Driver Behavior Event Detection for Manual Annotation by Clustering of the Driver Physiological Signals
Proceedings of Intelligent Transportation Systems (ITSC)
Naturalistic driving recordings are important for understanding the driver behavior. Driver behavior events of interest in these recordings, such as driver confusion and stress, are important for studying driver behavior and develop the next generation advanced driver assistant systems (ADASs). Unfortunately, such events are rare cases in the naturalistic driving data. Manual annotation is usually required to extract such events from a large data set. This study investigates the idea of using drivers’ physiological signals to help with the manual annotation process. The proposed framework uses the unsupervised cluster algorithm, density-based spatial clustering of applications with noise (DBSCAN), to cluster the physiological data into three classes: “Normal”, “Event” and “Noise”. We define three types of driver behavior events of interest in our real-world driving data, and evaluate the recall rate using the data classified in the “Event” cluster. High recall rate at 75% is achieved on average. We also evaluate the reduced effort for the annotator by estimating the viewing time compression rate, which is reduced by half when we set the fast forward rate in non “Event” segment to 5 times of normal speed.