Anthony D. McDonald

Steering in a random forest: ensemble learning for detecting drowsiness-related lane departures

Objective: The aim of this study was to design and evaluate an algorithm for detecting drowsiness-related lane departures by applying a random forest classifier to steering wheel angle data. Background: Although algorithms exist to detect and mitigate driver drowsiness, the high rate of false alarms and missed detection of drowsiness represent persistent challenges. Current algorithms use a variety of data sources, definitions of drowsiness, and machine learning approaches to detect drowsiness. Method: We develop a new approach for detecting drowsiness-related lane departures using steering wheel angle data that employ an ensemble definition of drowsiness and a random forest algorithm. Data collected from 72 participants driving the National Advanced Driving Simulator are used to train and evaluate the model. The model’s performance was assessed relative to a commonly used algorithm, percentage eye closure (PERCLOS). Results: The random forest steering algorithm had a higher classification accuracy and area under the receiver operating characteristic curve than PERCLOS and had comparable positive predictive value. The algorithm succeeds at identifying two key scenarios associated with the drowsiness detection task. These two scenarios consist of instances when drivers depart their lane because they fail to modulate their steering behavior according to the demands of the simulated road and instances when drivers correctly modulate their steering behavior according to the demands of the road. Conclusion: The random forest steering algorithm is a promising approach to detect driver drowsiness. The algorithm’s ties to consequences of drowsy driving suggest that it can be easily paired with mitigation systems.

Real-time detection of drowsiness related lane departures using steering wheel angle

Drowsy driving is a significant factor in many motor vehicle crashes in the United States and across the world. Efforts to reduce these crashes have developed numerous algorithms to detect both acute and chronic drowsiness. These algorithms employ behavioral and physiological data, and have used different machine learning techniques. This work proposes a new approach for detecting drowsiness related lane departures, which uses unfiltered steering wheel angle data and a random forest algorithm. Using a data set from the National Advanced Driving Simulator the algorithm was compared with a commonly used algorithm, PERCLOS and a simpler algorithm constructed from distribution parameters. The random forest algorithm had higher accuracy and Area Under the receiver operating characteristic Curve (AUC) than PERCLOS and had comparable positive predictive value. The results show that steering-angle can be used to predict drowsiness related lane-departures six seconds before they occur, and suggest that the random forest algorithm, when paired with an alert system, could significantly reduce vehicle crashes.

Text Mining to Decipher Free-Response Consumer Complaints: Insights From the NHTSA Vehicle Owner’s Complaint Database

Objective: This study applies text mining to extract clusters of vehicle problems and associated trends from free-response data in the National Highway Traffic Safety Administration’s vehicle owner’s complaint database. Background: As the automotive industry adopts new technologies, it is important to systematically assess the effect of these changes on traffic safety. Driving simulators, naturalistic driving data, and crash databases all contribute to a better understanding of how drivers respond to changing vehicle technology, but other approaches, such as automated analysis of incident reports, are needed. Method: Free-response data from incidents representing two severity levels (fatal incidents and incidents involving injury) were analyzed using a text mining approach: latent semantic analysis (LSA). LSA and hierarchical clustering identified clusters of complaints for each severity level, which were compared and analyzed across time. Results: Cluster analysis identified eight clusters of fatal incidents and six clusters of incidents involving injury. Comparisons showed that although the airbag clusters across the two severity levels have the same most frequent terms, the circumstances around the incidents differ. The time trends show clear increases in complaints surrounding the Ford/Firestone tire recall and the Toyota unintended acceleration recall. Increases in complaints may be partially driven by these recall announcements and the associated media attention. Conclusion: Text mining can reveal useful information from free-response databases that would otherwise be prohibitively time-consuming and difficult to summarize manually. Application: Text mining can extend human analysis capabilities for large free-response databases to support earlier detection of problems and more timely safety interventions.

The Language of Driving: Advantages and Applications of Symbolic Data Reduction for Analysis of Naturalistic Driving Data

Recent advances in onboard vehicle data recording devices have created an abundance of naturalistic driving data. The amount of data exceeds the resources available for analysis; this situation forces researchers to focus on analyses of critical events and to use simple heuristics to identify those events. Critical event analysis eliminates the context that can be critical in understanding driver behavior and can reduce the generalizability of the analysis. This work introduced a method of naturalistic driving data analysis that would allow researchers to examine entire data sets by reducing the sets by more than 90%. The method utilized a symbolic data reduction algorithm, symbolic aggregate approximation (SAX), which reduced time series data to a string of letters. SAX can be applied to any continuous measurement, and SAX output can be reintegrated into a data set to preserve categorical information. This work explored the application of SAX to speed and acceleration data from a naturalistic driving data set and demonstrated SAXs integration with other methods that could begin to tame the complexity of naturalistic data.

Variations on a theme: topic modeling of naturalistic driving data

This paper introduces Probabilistic Topic Modeling (PTM) as a promising approach to naturalistic driving data analyses. Naturalistic driving data present an unprecedented opportunity to understand driver behavior. Novel strategies are needed to achieve a more complete picture of these datasets than is provided by the local event-based analytic strategy that currently dominates the field. PTM is a text analysis method for uncovering word-based themes across documents. In this application, documents were represented by drives and words were created from speed and acceleration data using Symbolic Aggregate approximation (SAX). A twenty-topic Latent Dirichlet Allocation (LDA) topic model was developed using words from 10,705 documents (real-world drives) by 26 drivers. The resulting LDA model clustered the drives into meaningful topics. Topic membership probabilities were successfully used as features in subsequent analyses to differentiate between healthy drivers and those suffering from Obstructive Sleep Apnea.

Using kinematic driving data to detect sleep apnea treatment adherence

ABSTRACT People spend a significant amount of time behind the wheel of a car. Recent advances in data collection facilitate continuously monitoring this behavior. Previous work demonstrates the importance of this data in driving safety but does not extended beyond the driving domain. One potential extension of this data is to identify driver states related to health conditions such as obstructive sleep apnea (OSA). We collected driving data and medication adherence from a sample of 75 OSA patients over 3.5 months. We converted speed and acceleration behaviors to symbols using symbolic aggregate approximation and converted these symbols to pattern frequencies using a sliding window. The resulting frequency data was matched with treatment adherence information. A random forest model was trained on the data and evaluated using a held-aside test dataset. The random forest model detects lapses in treatment adherence. An assessment of variable importance suggests that the important patterns of driving in classification correspond to route decisions and patterns that may be associated with drowsy driving. The success of this approach suggests driving data may be valuable for evaluating new treatments, analyzing side effects of medications, and that the approach may benefit other drowsiness detection algorithms.

A contextual and temporal algorithm for driver drowsiness detection

This study designs and evaluates a contextual and temporal algorithm for detecting drowsiness-related lane. The algorithm uses steering angle, pedal input, vehicle speed and acceleration as input. Speed and acceleration are used to develop a real-time measure of driving context. These measures are integrated with a Dynamic Bayesian Network that considers the time dependencies in transitions between drowsiness and awake states. The Dynamic Bayesian Network algorithm is validated with data collected from 72 participants driving the National Advanced Driving Simulator. The algorithm has a significantly lower false positive rate than PERCLOS-the current gold standard-and baseline, non-contextual, algorithms under design parameters that prioritize drowsiness detection. Under these parameters, the algorithm reduces false positive rate in highway and rural environments, which are typically problematic for vehicle-based detection algorithms. This algorithm is a promising new approach to driver impairment detection and suggests contextual factors should be considered in subsequent algorithm development processes. It may be combined with comprehensive mitigation methods to improve driving safety.

Highway Healthcare: How Naturalistic Driving Data Index Adherence to CPAP Therapy in Obstructive Sleep Apnea

Drowsy driving is a major factor in many vehicle crashes around the world. Sleep disorders, such as ob-structive sleep apnea (OSA), underpin many of these crashes. Continuous positive airway pressure (CPAP) therapy is an effective treatment for sleep apnea but it requires consistent use and is often rejected by OSA patients. Rejection of CPAP treatment creates a dangerous on-road environment for both OSA sufferers and the general public. Algorithms capable of detecting CPAP use and its effects on driving are integral to iden-tifying and mitigating this danger. This work uses naturalistic kinematic driving data to develop an algo-rithm which can detect nightly CPAP abstinence and adequate CPAP use. Speed and lateral acceleration data were collected using a data recorder in participant’s primary vehicle and CPAP data were collected by downloading adherence data from participant CPAP machines. The speed and acceleration data were re-duced to a set of symbols using Symbolic Aggregate approximation (SAX) time-series analysis. The sym-bols were converted into a sequence frequency dataset using sliding windows of size 1 to 10 s with a 1 Hz sampling rate. A Random Forest classifier was trained on the data to create a classification algorithm. On a held aside testing set, the Random Forest algorithm correctly identified 71% of the instances and had an area under the receiver operating characteristic curve of 0.76. The variable importance of the algorithm suggested that kinematic patterns associated with common drowsy driver crash types were key features in the algorithm’s prediction performance.