Driver Attention Level Estimation Of Non-driving Activity Based On Computer Vision

In the development of modern automobile industry,vehicle is not only a tool for daily travel anymore,but a mobile and interconnected space with various possibility in the future,and it will meet the requirements of safety,intelligence,efficiency and entertainment better and better.Vehicle intelligence,interconnection,electric and sharing is the trend according to the interest of consumer,and intelligent vehicle will play a critical role in this revolution,which refers to complex environment perception,data transferring and computing,decision and execution.L3 intelligent vehicle will be a feasible transition before the key technology are fully mature and generally commercial,and it can also accumulate massive amounts of data and experience for higher intelligence,and accelerate the industry application process of automation as a result.For L3 intelligent vehicle,when the vehicle is under the automatic driving mode,the driver may participate in non-driving activities such as making a phone call,playing a tablet.But when the system meets abnormal complex environment or sub-system lose function,the driver is supposed to taking over and giving up the current non-driving activities.Therefore,how to monitor the non-driving activity attention level of driver when the system is under automation mode is important.At present,the research on the driver attention level mainly focuses on traditional driving monitoring which includes distraction and fatigue.However,there are few research on the L3 intelligent non-driving activity,so this paper spotlight this scenario and conducts research on the following aspects:(1)This topic is based on the L3 intelligent driving scenario,in which the driver participates in the non-driving activity when the vehicle driving automatically.And giving up the current non-driving activity to take over the control when automation function failed.So the attention level towards non-driving activity is crucial to this take over process such as the takeover signal design and take over strategy design.(2)Establishing the head movement estimation system based on computer vision for the L3 intelligent vehicle scenario,applying pixel intensity comparison +gentle boost algorithm to find out the face region,and then detecting 68 facial landmark based on Constrained Local Neural Fields method,solving the parameter between 3D world coordinate and 2D pixel image coordinate according to EPNP algorithm in the end.One common webcam is enough to capture image,and computing demand is relatively low.What's more,the detection speed is fast and detection accuracy is acceptable,so it's practical for this application.(3)Typical non-driving activity playing tablet is focused for detailed analysis,and exploring the relationship of head movement and tablet movement in different attention level.For this purpose,a detailed experiment planning including experiment content and supervision is designed to collect movement data of head and tablet.Then applying time-varying cross-correlation analysis to find out the extract feature,and then constructing attention level estimation model based on support vector machine.After that,using sensitivity analysis method to study the influence of parameters on model accuracy.Finally,the model is verified through dynamic experiment data and shows a good result on attention level estimation.