Identifying Fine-grained Driving Style With GPS Trajectory Data

Driving behaviour understanding plays a vital role that can improve transportation safety and promote the development of Intelligent Transportation Systems(ITS).Driving style is an abstract yet important description and interpretation of driving behaviour.As a long-standing research topic in driving behaviour analysis,driving style identification is non-trivial.With the rapid development of mobile internet and internet of things technologies during recent years,the massive perceptual data generation provides important data resources for driving behaviour analysis.Almost all previous studies about driving style identification emphasize the research on the granularity of an entire trip or a driver,probably leading to failing to capture the time-varying and complex driving behaviours.Inspired by the fact that an aggressive driver may drive safely at some time.Based on the pervasive and easily-collected GPS trajectory data,this thesis proposes the following three innovative solutions to identify fined-grained driving style effectively.First,this thesis proposes a driving style identification model based on graph convolutional networks.To be more specific,first and foremost,based on the low quality and rough GPS trajectory data,the fine-grained driving behaviour expression is conducted to construct the graph data tightly related to driving behavior and style from multiple perspectives.Then,graph convolutional networks are designed to effectively integrate complementary information provided by multi-view graph data.Finally,the proposed model is evaluated extensively based on real-life taxi trajectory dataset collected,and the experimental results demonstrate that it outperforms other baselines.Second,this thesis proposes a driving style identification model based on graph attention networks.Similarly,based on the multi-view graph data built by the fine-grained driving behaviour expression,the graph attention networks are designed to extract high-level and discriminable features regarding complex driving behaviours.Then,the proposed model is evaluated extensively based on real-life taxi trajectory dataset collected.Compared to the identification model based on graph convolutional networks,the experimental results show that it achieves better identification performance,and the training time of the model is acceptable.Third,to reduce the burden of time-consuming manual labeling and achieve the better identification performance simultaneously,this thesis further proposes a driving style identification framework based on the semi-supervised learning.More specifically,based on any of the above-proposed identification models,a simple yet effective semi-supervised learning method called Pseudo-Label is adopted to make use of the unlabelled data sample and to further improve the identification performance.Then,the proposed framework is evaluated extensively based on real-life taxi trajectory dataset collected,and the experimental results demonstrate its effectiveness.