Transfer-learning-enabled driver behavior model adaptation towards cognitive autonomous driving

点击翻译按钮获取中文摘要

There are about 1.2 million road fatalities worldwide every year in which driver behavior or driver error is the most critical causal factor, and it is widely expected that increasing vehicle autonomy will drastically reduce driver error. Governments and industry around the world are recognizing the far-reaching impacts that will eventually be delivered from vehicle automation. A Canadian Senate Committee on Transport and Communications report, published in Jan 2018, gave 16 recommendations to help prepare Canadians for autonomous vehicles, highlighting that Canada must plan better for the impacts of this emerging technology. The coming decades will bring mixed traffic flows, with human-driven and automated vehicles sharing the road. These complex interactions between human drivers and autonomous driving systems (within the same vehicle or among different vehicles) have significant potential to induce human driver errors and/or autonomous driving system decision errors. Such challenges illustrate the emerging need for further enhanced understanding and modeling of human driver behaviors, though considerable efforts have been made in this field in the past few decades. The long-term goal of this research program is to enhance the science and understanding in human driver behaviors towards cognitive autonomous driving, under various driving scenarios, which include normal driving scenarios as well as critical scenarios (or corner cases). Towards achieving this long-term goal, this discovery grant focuses on development and validation of a new transfer learning framework for driver behavior model adaptation. Apart from the available public dataset, new data will be collected using driving simulator and real vehicles, for model training and validation. The proposed research program along with the research outcomes will contribute considerably to the enhanced understanding and modeling of human driver behaviors (and also human behaviors in daily activities), which is also very valuable for the emerging development of cognitive autonomous driving, where the interaction-awareness capacity of autonomous vehicles is needed with surrounding vehicles as well as the operator inside. As an additional valuable outcome, the datasets collected within this program will be published for public research purposes. This research program will have a very valuable impact in automated-driving-oriented companies or startups in Canada. Since Canadian companies will be the primary receptors of these technologies, the research efforts in this program will contribute to advancing future-generation Canadian-made automated driving vehicles with enhanced intelligence and safety. Furthermore, through disseminations, the outcomes of this research program will assist in boosting the public acceptance, offer recommendations related to human driver behaviors for policy makers, and accelerate the penetration of automated driving vehicles in Canada and worldwide.