|This session presents methods for examining the behavior of distracted driving, predicting crashes and their severity, detecting aggressive and unsafe driving behavior, and sensing the environment for collision avoidance in advanced driver assistance systems. The applications demonstrate the use of machine learning tools such as random forests, convolutional neural networks, deep belief neural networks, clustering techniques, and others.|
Beyond Grand Theft Auto V for Training, Testing, and Enhancing Deep Learning in Self-Driving Cars
Mark Martinez, Princeton UniversityShow Abstract
Chawin Sitawarin, Princeton University
Kevin Finch, Princeton University
Lennart Meincke, NordAkademie
Alexander Yablonski, Princeton University
Alain Kornhauser, Princeton University
As an initial assessment, over 480,000 labeled virtual images of normal highway driving were readily generated in Grand Theft Auto V's virtual environment. Using these images, a CNN was trained to detect following distance to cars/objects ahead, lane markings, and driving angle (angular heading relative to lane centerline): all variables necessary for basic autonomous driving. Encouraging results were obtained when tested on over 50,000 labeled virtual images from substantially different GTA V driving environments. This initial assessment begins to define both the range and scope of the labeled images needed for training as well as the range and scope of labeled images needed for testing a the definition of boundaries and limitations of trained networks. It is the efficacy and flexibility of a "GTA V"-like virtual environment that is expected to provide an efficient well-defined foundation for the training and testing testing of Convolutional Neural Networks for safe driving. Additionally, described is the Princeton Virtual Environment (PVE) for the training, testing and enhancement of safe driving AI, which is being developed using the video-game modeling engine Unity. PVE is being developed to recreate rare but critical corner cases that can be used in re-training and enhancing machine learning models and understanding the limitations of current self driving models. The Florida Tesla crash is being used as an initial reference.
Classification of Distracted Driving Based on Visual Features and Behavior Data Using a Random Forest Method
Xiaohua Zhao, Beijing University of TechnologyShow Abstract
Ying Yao, Beijing University of Technology
Hongji Du, Beijing University of Technology
Yunlong Zhang, Texas A&M University
This research uses data collected from a driving simulator and an eye tracking system to explore the relationship between a driver’s visual features and driving behaviors of distracted driving, and a Random Forest (RF) method is developed to classify driving behaviors and improve the accuracy of detecting distracted driving. Drivers are required to complete four distraction tasks while they followed simulated vehicles in the experiment. In data analysis, the features of distracted driving behaviors are first described, while the visual data are classified into three distraction levels based on the AttenD algorithm. Based on this collected data, this paper detail the relationship between visual features and driving behavior. Significant differences are discovered between different distraction tasks and distraction levels. Additionally, driving behavior data is used to build a RF model to classify distracted driving into three levels. Results demonstrate that this model is feasible to capture the classification of distraction and its accuracy for each distraction task is over 90%. AUCs calculated through Error-Correcting Output Codes are mainly around 0.9, indicating good reliability. With this classification method, distraction levels could be classified with vehicle operation characteristics. The model established by this method could can detect distractions in actual driving through the detection of driving behavior without the need of eye tracking systems.
An Improved Deep Belief Network Model for Road Safety Analyses
Guangyuan Pan, University of WaterlooShow Abstract
Liping Fu, University of Waterloo
Lalita Thakali, University of Waterloo
Matthew Muresan, University of Waterloo
Ming Yu, University of Waterloo
Crash prediction is a critical component of road safety analyses. A widely adopted approach to crash prediction is application of regression based techniques. The underlying calibration process is often time-consuming, requiring significant domain knowledge and expertise and cannot be easily automated. This paper introduces a new machine learning (ML) based approach as an alternative to the traditional techniques. The proposed ML model is called regularized deep belief network (DBN), which is a deep neural network with two training steps: it is first trained using an unsupervised learning algorithm and then fine-tuned by initializing a Bayesian neural network with the trained weights from the first step. The resulting model is expected to have improved prediction power and reduced need for the time-consuming human intervention. In this paper, we attempt to demonstrate the potential of this new model for crash prediction through two case studies including a collision data set from 800 km stretch of Highway 401 and other highways in Ontario, Canada. Our intention is to show the performance of this ML approach in comparison to various traditional models including negative binomial (NB) model, kernel regression (KR), and Bayesian neural network (Bayesian NN). We also attempt to address other related issues such as effect of training data size and training parameters.
Mobile Sensing and Machine Learning for Identifying Driving Safety Profiles
Eleni Mantouka, National Technical University of Athens (NTUA)Show Abstract
Emmanouil Barmpounakis, National Technical University of Athens (NTUA)
Eleni Vlahogianni, National Technical University of Athens (NTUA)
A large number of drivers with different driving characteristics co-exist on the road network. Assessing a person’s driving profile and detecting aggressive and unsafe driving behavior is essential to enhance road safety, reduce fuel consumption and - at a macroscopic level - tackle congestion. Nowadays, driving data can be massively collected via sensors embedded in mobile phones, avoiding the expensive and inefficient solutions of in-vehicle devices. In this paper, these data are used to detect unsafe driving styles based on two-stage clustering approach and using information on harsh events occurrence, acceleration profile, mobile usage and speeding. First, an initial clustering was performed in order to separate aggressive from non–aggressive trips. Subsequently, to distinguish "normal" trips from unsafe trips, a second level clustering was performed. In this way, trips have been categorized into six distinct groups with increasing importance with respect to safety. The further analysis of drivers in relation to the grouping of their trips showed that drivers cannot maintain a stable driving profile through time, but exhibit a strong volatile behavior per trip. Finally, a discussion is provided on the implications of the main findings in research and practice.
Adaptable Advanced Driver Assistance Systems (ADASs)
Anthony Ohazulike, Hitachi Europe SASShow Abstract
Fully autonomous cars will be realized earlier than expected. Artificial intelligence (AI) will play a vital role in the development of self-driving cars. Challenges of self-driving cars include the ability of the vehicle to sense its environment and avoid collisions with objects, determine what action to take if unexpected events occur, and how to make the user feel comfortable in the self-driving car. One way to tackle the later challenge is by adapting the driving style of the self-driving car to that of the driver. Advanced Driver Assistance Systems (ADAS) is a system design to help relieve driver of some of the driving tasks. With respect to ADAS, this paper describes a real-time AI-based approach that continuously learns driving behavior, and adapts the Driver Assistance System to the current driving style of the driver. Such AI adaptation ensures that ADAS, while relieving the driver of certain tasks, does so in a way very natural to this driver. A field test confirmed that the AI model described in this paper achieved an accuracy level of 98%. In addition, the paper provides a model that can anticipate driving maneuver. Data collected on different categories of drivers, can be used as bases for creating driving profiles for first fully autonomous cars.