Regularized Deep Learning And Its Application In Driving Safety Risk Assessment

The statistics reports released by the World Health Organization(WHO)show that motor vehicle road traffic accidents cause about 1.35 million deaths and more than 50 million injuries every year.Moreover,due to China's huge car parc,the number of deaths per year in China's road traffic accidents accounts for 22.6%of the world's total,ranking first in the world.In addition,a large number of studies have shown that driving behaviors have a strong correlation with road traffic safety risks,and more than 90%of road traffic accidents are caused by drivers' dangerous driving behavior.Therefore,in order to further improve the overall safety of China's road network and reduce the occurrence of motor vehicle traffic accidents,monitoring,standardizing and controlling the driving behavior of individual drivers is an effective solution.With the rapid development of Intelligent Transportation Systems(ITS),high-resolution multi-dimensional data such as "people,vehicles,roads,and environment"can be real-time acquired through smart devices,which provide a vast array of underlying data sources for in-depth analysis and understanding of driving behavior and driving safety risks.In addition,deep learning,because of its powerful abstract feature representation performance,can be used as a powerful technical method to deeply explore the temporal and spatial rules and behavioral risks implied by massive driving behavior data.This thesis firstly studies the key technical problems of deep learning in feature representation and regularization.Based on the improved deep learning technology,the massive multi-dimensional real-time data such as "people,vehicles and roads" is fully integrated to establish an overall view of driving events,and to more comprehensively explore the temporal and spatial rules implied by driving behavior.In this way,we can deeply understand the corresponding driving behavior patterns and the behavioral risks,and then analyze the relationship between driving behavior and driving safety risks from a more scientific and complete perspective.This thesis provides theoretical foundation and technical solutions for areas such as improving the traditional deep learning performance,standardizing driving behavior for improving road traffic safety,driving behavior-based auto insurance premiums,and advanced driver assistance systems.The main research contents of this thesis are as follows:(1)This thesis proposes a driving behavior identification method based on cross-covariance constraint deep autoencoder.Firstly,existing feature representation algorithms based on DAE tend to produce duplicative encoding and decoding filters,which leads to feature redundancy and overfitting.This thesis using the cross-covariance to regularize the feature weight vector to construct a new objective function to eliminate feature redundancy and reduce overfitting.Further,Existing driving behavior identification methods have the disadvantages that the size of the sliding time window is too large causes the model training and prediction time to be long,and the feature extraction is relatively subjective and it is difficult to completely capture the spatiotemporal feature information of the driving behavior.This thesis constructs a driving behavior identification method based on XCov regularized Nonnegativity-Constrained Autoencoder(XCov-NCAE),which can fully integrate multi-dimensional information of the driving behavior data and automatically extract more representative hidden features of the driving behavior,thus achieving precise depiction of driving behaviors of different styles and patterns.The effectiveness of the proposed method is verified based on the large-scale natural driving behavior dataset.(2)This thesis proposes a driver distraction monitoring method based on deep relevant feature representation architectures.Firstly,to solve the problem of hidden feature information loss caused by Dropout,this thesis proposes a regularization strategy called DropMI(Mutual Information-Based Dropout,DropMI),which introduces Mutual information(MI)to evaluate the importance of hidden layer neural units to target representation,and constructs a new binary mask matrix based on the sorting distribution of MI,and then developing a dynamic DropMI strategy to improve the performance of relevant feature representation of deep networks.Furthermore,to solve the problems of the existing research methods of driver distraction monitoring,such as single acquisition index,strong subjective feature extraction and lack of recognition of different types of distraction statuses,this thesis constructs the deep relevant feature representation architectures based on DropMI,which fully integrates multi-source driver status informations(including ECG,EEG,vision,driving vehicle status,etc.)and accurately monitors different types of distraction status of the driver(including cognitive distraction,emotional distraction,and sensorimotor distraction).The effectiveness of the proposed method is verified based on the multimodal driver distraction dataset.(3)This thesis proposes a novel cost-sensitive L1/L2-nonnegativity-constrained deep autoencoder network for driving safety risk prediction.To solve the problems of the existing research methods of the feature extraction is relatively subjective and does not fully consider the driving behavior and related factors of near-crash,and the class imbalance caused by the small number of the crash event,this thesis first establishes a spectral model of driving safety risk and driving critical events(normal,near-crash and crash)to define the correspondence between different driving critical events and driving safety risk levels,and then the fully integrate the time and space factors related to driving critical events,and construct a deep L1/L2-Nonnegativity-Constrained Autoencoder(L1/L2-NCAE)for unsupervised automatic extract hidden features of driving behaviors of different risk levels.In addition,to solve the class imbalance,this thesis proposes the L1/L2-Nonnegativity-Constrained Focal Loss(L1/L2-NCFL)classifier to predict different levels of driving safety risks.The effectiveness of the proposed method is verified based on the 100-CAR Naturalistic Driving Studies(NDS)dataset.