Adaptive Control Of Advanced Driver Assistance Systems Based On Personalized Driver Models

Accurate individual driver model is one of the keys to develop an intelligent vehicle systems capable of adapting to this driver’s demands.However,dynamically stochastic char-acteristics of driver behaviors bring great challenges in achieving such advanced driver as-sistance systems(ADAS).The dynamic and stochastic processes of driver behavior was in-vestigated and applied to ADAS,including:The dynamic and stochastic processes of driver behavior was modeled based on modern statistical learning theory and combined with lateral and longitudinal driving assistance sys-tems.By considering the attributes of non-Gaussian driving data with bounded supports,a bounded generalized GMM-HMM(BGGMM-HMM)personalized driver model was devel-oped.And a model parameter estimation algorithm with Monte Carlo sampling theory was developed.A series of systematic validations and analysis in specific driving scenarios were also conducted using real vehicles.More specifically,(1)First,for modeling dynamic and stochastic processes of driver behaviors,the state-of-the-art driver models were reviewed and their pros and cons were shown and analyzed.Considering stochastic and dynamic characteristics of driver behavior,this thesis developed a GMM-HMM personalized driver model based on the modern statistical learning theory,which can capture the dynamically stochastic characteristics of driver behavior in terms of decision-making processes.In addition,the effectiveness of the proposed model was primar-ily demonstrated in car-following scenarios.(2)Second,for modeling and prediction of driver’s longitudinal control behavior,a perception-decision-action personalized driver model was developed based on GMM-HMM.This proposed model overcomes the limitations on traditional approaches such as prediction delay and light-dependent and is able to predict driver’s braking behavior in car-following scenarios.The comparison results show its advantages and effectiveness.(3)Third,for modeling and predicting human driver’s lateral departure behavior in a dynamic and stochastic environment,a GMM-HMM personalized driver model was built and learned based on the naturalistic driving data collected from this person.An iterative algorithm to predict human driver’s forthcoming behavior based on the learned personalized driver model was proposed.An associated lane departure warning algorithm was proposed to reduce the false alarm rate,thus improving the acceptability of lane departure warning systems for drivers.(4)Fourth,in order to overcome the limitations of GMM-HMM on describing the non-Gaussian distribution data with bounded support,a generalized GMM-HMM model with con-sidering data bounded support,i.e.,BGGMM-HMM,was proposed by introducing a bounded function in measure space and a model shape parameter.In addition,an associated iterative algorithm was also developed to estimate model parameters based on the Monte Carlo sam-pling theory.(5)Fifth,in order to model and predict vehicle acceleration when following a car,the developed BGGMM-HMM model was employed with considering the non-Gaussian dis-tribution with bounded support of driving data.Other three modern approaches,including SMM-HMM,GMM-HMM,and GGMM-HMM were also developed to show its effective-ness in terms of high accuracy,strong robustness,and great generalization capability.