Abstract

Driver distraction is a leading cause of crashes. The introduction of in-vehicle technology in the last decades has added support to the driving task. However, in-vehicle technologies and handheld electronic devices may also be a threat to driver safety due to information overload and distraction. Adaptive in-vehicle information systems may be a solution to this problem. Adaptive systems could aid the driver in obtaining information from the device (by reducing information density) or prevent distraction by not presenting or delaying information when the driver’s workload is high. In this paper, we describe an on-the-road evaluation of a real-time driver workload estimator that makes use of geo-specific information. The results demonstrate the relative validity of our experimental methods and show the potential for using location-based adaptive in-vehicle systems.

Abstract

International audience; Driver distraction is one of the leading causes of fatal car accidents. Driver distraction is any task that diverts the driver attention from the primary task of driving and increases the driver’s cognitive load. Detecting potentially dangerous driving situations or automating some repetitive tasks, using Advanced Driver Assistance Systems (ADAS), and using autonomous vehicles to reduce human errors while driving are two suggested solutions to diminish driver distraction. These solutions have some advantages, but they suffer from their inherent inability to detect all potentially dangerous driving situations. Besides, autonomous vehicles and ADAS depend on sensors. As a result, their accuracy diminishes significantly in adverse conditions. Analyzing driver behavior using machine learning methods and estimating the distraction level of drivers can be used to detect potentially hazardous situations and warn the drivers. We conducted an experiment in eight different driving scenarios and collected a large dataset from driving data and driver related data. We chose Long Short Term Memory (LSTM) as our machine learning method. We built and trained a stacked LSTM network to estimate the driver status using a sequence of driving data vectors. Each driving data vector has 10 driving related features. We can accurately estimate the driver status with no external devices and only using cars Can-Bus data. Document type: Conference object