Enhancing the Architecture of Context-Aware Driver Assistance Systems by Incorporating Insights from Naturalistic Driving Data

Abstract

Driving assistance systems (DASs) have received a great deal of attention in the past decades as an active and effective collision countermeasure. DASs potential benefits will be attained by enhancing the systems’ awareness regarding the dynamic driving context including the change in the driver behavior, vehicle status, and surrounding environment. Therefore, context-aware systems were proposed to improve the adaptability of the DASs to that dynamic driving context. This dissertation proposes a new aspect to the architecture of the context-aware DASs by introducing a context identification layer to the reasoning subsystem. The main goal of this layer is to accommodate changes in driver behavior due to the surrounding environment and to customize the system to the needs of each individual driver. The proposed layer is designed based on key insights that were drawn from naturalistic driving studies based on actual driving behavior. Data from 64 drivers in a Naturalistic Driving Study (NDS) was selected as a measurement of driver behavior because it provides network-wide data collection without restrictions or instructions to the driver on the routes that should be used or the roadway elements (e.g., curves or two-way two-lane roads), making the data collected closest to reality. The data was processed, and events of interest were extracted, mapped using ArcGIS, and analyzed to differentiate between the intersection- and segment-related events. The intersection-related events were identified according to the intersection influence area, which was estimated based on the stopping sight distance and the speed limit. Several behavioral measures were extracted for each event of interest, including following distance, relative speed, headway between the host vehicle and the leading vehicle, acceleration, time-to-collision, and jerk of the host vehicle.Based on the insights observed, the proposed context identification layer contains two algorithms that work in sequence: the infrastructure detection algorithm and the driver classification algorithm, respectively. The infrastructure detection algorithm aims at identifying intersection-related driving where the driver adjusts his or her behavior due to the presence of an intersection ahead. This algorithm was developed by training a deep neural network for each driver to accommodate each driver needs. The output of this was then used to initiate the processing of the driver classification algorithm. The driver classification algorithm categorizes drivers into cautious, normal, and aggressive drivers using the first algorithm (i.e., near intersections or on segments) and combining behavioral measures using unsupervised machine learning techniques including principal component analysis and K-means. The provided key insights into driver behavior included, but was not limited to: (i) the quantification of driving behavior in proximity to intersections and on segments in the form of probability density functions, (ii) there were significant differences between both behaviors in terms of the measures used, (iii) the proposed definition of the intersection influence area was plausible in terms of distinguishing between the segment- and intersection-related driving behavior, (iv) acceleration was the highest among the behavioral measures influenced by proximity to an intersection, and (v) the driver classification, which ignores the driver’s relative location to intersections, were more likely to misclassify drivers as aggressive when they were in high intersection density areas such as downtown cores.