Online Driver Distraction Detection Using Long Short-Term Memory

Abstract

Abstract-Lane-keeping assistance systems for vehicles may be more acceptable to users if the assistance was adaptive to the driver's state. To adapt systems in this way, a method for detection of driver distraction is needed. Thus, we propose a novel technique for on-line detection of driver's distraction, modeling the long-range temporal context of driving and head tracking data. We show that Long Short-Term Memory (LSTM) recurrent neural networks enable a reliable, subject-independent detection of inattention with an accuracy of up to 96.6 %. Thereby our LSTM framework significantly outperforms conventional approaches such as Support Vector Machines.

Investigating the Feasibility of Vehicle Telemetry Data as a Means of Predicting Driver Workload

“…This is likely due to the higher responsiveness of physiological measures to workload, and performance measures extracted from vehicle telemetry data are often used only as secondary inputs (Wollmer et al, 2011). Vehicle telemetry data includes measurements from all devices in the vehicle, and can be recorded via the Controller Area Network (CAN)-bus.…”

“…Vehicle telemetry data includes measurements from all devices in the vehicle, and can be recorded via the Controller Area Network (CAN)-bus. Common measures used for driver workload monitoring include features extracted from the steering wheel, vehicle speed, and pedal positions (Wollmer et al, 2011;Mehler et al, 2012). The mean or standard deviation (STD) is often extracted from signals over whole distraction or normal driving periods, and are often minutes long.…”

Redundant Feature Selection for Telemetry Data

“…Based on these inferences, the agent can determine whether or not to present the driver with new information that might unnecessarily add to their workload. Traditionally, such agents have monitored physiological signals such as heart rate or skin conductance [2,16]. However, such approaches are not practical for everyday use, as drivers cannot be expected to attach electrodes to themselves before driving.…”

“…These include statistical measures such as the mean, minimum or maximum, as well as derivatives, integrals and spectral measures. In [16], Wollmer et al extract a total of 55 statistical features over temporal windows of 3 seconds from 18 signals including steering wheel angle, throttle position and speed, and driver head position. This provides a total of 990 features for assessing online driver distraction.…”

“…In many cases the features themselves are extracted from multiple signal data such as vehicle telemetry [16], medical sensors [14], weather forecasting [10], and financial time-series analysis [13]. When features are extracted from signal data, it is possible for feature redundancy to be either between features extracted from the same signal (intra-signal), or between features extracted from different signals (inter-signal).…”