Abstract
Keywords: Driver Assistance System, Driver Behaviour, Driving Database, Statistical Machine Learning, Active Safety
Recent advances in practical adaptive driver assistance systems and sensing technology have led to a detailed study of individual driver behaviour. The design of cooperative control among the currently developed active safety devices, which fits the driver behaviour/intention and ongoing traffic situation, is required. To realize such systems, an extensive study of individual driver behaviour model is necessary. This paper describes the modelling of naturalistic driving behaviour in real-world traffic situation, based on driving data collected by using the experimental car equipped with the continuous sensing drive recorder. The driving route includes several types of roadways in urban area of Tokyo. The continuous sensing drive recorder includes the information of road environment (GPS data and headway distance), driver operation (steering, pedal, winker), and vehicle dynamics data (speed, acceleration, yaw rate, etc.). The video images of front scenery, rear scenery, driver's face and driver's foot operation are captured in synchronization with the sensor data. The real-time driving state recognition algorithm, in longitudinal vehicle operation, is described as one of the core technology of individual adaptation technique. The driver model in urban area is assumed to be in the form of state flow diagram. In this paper, the longitudinal driving sequence is classified into five categories with symbols corresponding to driving states (modes): car following, braking (when there is a preceding vehicle, obstacle), free following (independent driving), deceleration (when there is no preceding vehicle) and stopping. In this paper, Boosting Sequential Labelling Method, which describes the relationship between the sensor data of drive recorder and the symbolized driving state as the conditional probability feature, is employed to train the driver-vehicle-environment model in the manner of datadriven method for recognizing driving manoeuvres in real time. The conditional probability of driving sequential label with respect to drive recorder sensor data can be computed based on Logistic Regression Model. The combination of optimized weak classifiers, described as ifthen rules, represents the classification of each driving state. In the full paper, the accuracy of the estimation method, evaluated as F-value, is investigated by comparing the estimated label with the ground truth of the driving state labels.