Research On Multi-modal Fusion Evaluation Model Of L3 Intelligent Vehicle Driver Takeover Ability

The driving right of L2 assisted driving system has not been transferred,and only the driver’s fatigue and distraction need to be monitored.Unlike L2 assisted driving system,L3 system requires the transition from driving right to driver in case of emergency.Therefore,the evaluation of driver takeover ability has become a new direction of driver condition monitoring research.Takeover capability assessment based on driver behavior features has attracted much attention because it can make a direct judgment on whether the driver can drive safely after takeover.However,existing research for over more capability assessment of vehicle braking reaction time and takeover behavior as evaluation index,due to the driving power is transferred to the class index for reserve decision time is less,this paper proposes a fusion to takeover the former physiological signals,facial expressions such as modal features of the method,starting from the features of the driver to takeover the former,Assess whether the driver has the ability to drive safely after takeover.Firstly,a scene experiment is designed to collect experimental data.In order to use the features of the driver before takeover to predict the quality of the driver takeover,22 experienced drivers are recruited to participate in the driving simulation experiment,complete the takeover task when the vehicle speed in the takeover scene is110km/h and the collision time TTC is 6S,and record the physiological signal,takeover reaction time and facial expression of the driver before takeover as the raw data,The driving condition after takeover shall be used as the basis for the evaluation of takeover ability.Then,the multi-modal original signal feature extraction,physiological signal chooses ECG signal,by filtering out noise power frequency and remove outliers,and then refer to the driver condition monitoring of physiological research,extraction time and frequency domain features;The reaction time of takeover could be reached by calculating the time interval between the sending of takeover signal and the driver pressing the takeover button.Facial features based on the pilot experimental video frame,and with the help of CNN on semantic segmentation exclude the interference of background factors,and then use based on facial expressions open data set training model expression classification.To extract the modal features of correlation analysis,from the perspective of statistical validation of a single species features is difficult to rule out individual differences,namely multi-modal information fusion method of necessity.Finally,the combination of feature level fusion and part decision level fusion is used for classification and prediction,from the perspective of accuracy and ROC curve to evaluate model.The results showed that: Prediction model when input features in multi-modal information fusion,the estimating accuracy and generalization ability of takeover ability is higher than single species features,and has good ability to prevent a fitting.Sorting different types of features based on XGBoost,the results show that in the case of multi-modal information fusion according to the features of source,all kinds of features of the comprehensive weight of equilibrium,and to takeover the quality of prediction accuracy than the multi-modal information fusion for promotion.