A Context Aware Classification System for Monitoring Driver’s Distraction Levels

Date

2021

Advisors

Journal Title

Journal ISSN

ISSN

DOI

Volume Title

Publisher

De Montfort University

Type

Thesis or dissertation

Peer reviewed

Abstract

Understanding the safety measures regarding developing self-driving futuristic cars is a concern for decision-makers, civil society, consumer groups, and manufacturers. The researchers are trying to thoroughly test and simulate various driving contexts to make these cars fully secure for road users. Including the vehicle’ surroundings offer an ideal way to monitor context-aware situations and incorporate the various hazards. In this regard, different studies have analysed drivers’ behaviour under different case scenarios and scrutinised the external environment to obtain a holistic view of vehicles and the environment. Studies showed that the primary cause of road accidents is driver distraction, and there is a thin line that separates the transition from careless to dangerous. While there has been a significant improvement in advanced driver assistance systems, the current measures neither detect the severity of the distraction levels nor the context-aware, which can aid in preventing accidents. Also, no compact study provides a complete model for transitioning control from the driver to the vehicle when a high degree of distraction is detected. The current study proposes a context-aware severity model to detect safety issues related to driver’s distractions, considering the physiological attributes, the activities, and context-aware situations such as environment and vehicle. Thereby, a novel three-phase Fast Recurrent Convolutional Neural Network (Fast-RCNN) architecture addresses the physiological attributes. Secondly, a novel two-tier FRCNN-LSTM framework is devised to classify the severity of driver distraction. Thirdly, a Dynamic Bayesian Network (DBN) for the prediction of driver distraction. The study further proposes the Multiclass Driver Distraction Risk Assessment (MDDRA) model, which can be adopted in a context-aware driving distraction scenario. Finally, a 3-way hybrid CNN-DBN-LSTM multiclass degree of driver distraction according to severity level is developed. In addition, a Hidden Markov Driver Distraction Severity Model (HMDDSM) for the transitioning of control from the driver to the vehicle when a high degree of distraction is detected. This work tests and evaluates the proposed models using the multi-view TeleFOT naturalistic driving study data and the American University of Cairo dataset (AUCD). The evaluation of the developed models was performed using cross-correlation, hybrid cross-correlations, K-Folds validation. The results show that the technique effectively learns and adopts safety measures related to the severity of driver distraction. In addition, the results also show that while a driver is in a dangerous distraction state, the control can be shifted from driver to vehicle in a systematic manner.

Description

Keywords

Citation

Rights

Research Institute

Collections