Research On The Lane-changing Style Classification Method Based On Driving Behavior Primitives

With the development of intelligent automobile,advanced driving assitance systems have come into the public and have been widely used,such as Lane Keeping Assistance system and Lane Departure Warning system,which play a significant role in ensuring driving safety,improving driving experience and improving traffic efficiency.However,the current driving assistance systems lacks of deep consideration for the diversity of drivers’ driving styles,and it is difficult to meet the individual needs of different drivers.Most of the existing studies on the classification of driving styles are based on the explicit features of driving data,but these data contain rich implicit semantic features,which cover the driver’s handling characteristics.If such implicit features can be mined from driving data,complex driving behaviors can be decomposed into simple,small and primitive segments,which will be of great significance for understanding driver’s driving style and driving behaviors.Thus,in this paper,this segment is defined as the driving behavior primitive and based on it,the classification of lane-changing style is studied.Slice processing of lanechanging can be realized by extracting driving behavior primitive,so as to deeply understand lane-changing behavior and characteristics of drivers.Therefore,the classification of drivers’ driving style can be studied from a new point.Such a deep-seated semantic analysis of driving behavior is helpful to improve the anthropomorphic level of intelligent automobile decision making,and make more human-oriented control behaviors,thus contributing to the development of driving assistance systems and human-like autonomous vehicles.At the same time,it can make foundation for the formulation of traffic laws and the evaluation of vehicle insurance.Supported on the National Natural Science Foundation of China(52172386),and National Natural Science Foundation of China(51775235),this paper has been carried out the research on the lane-changing style classification method based on driving behavior primitives.The lane-changing behavior has been deeply analyzed,and then the screening rules of lanechanging trajctory have been formulated.Based on the rules,the original trajctory of lanechanging has been extracted from the natural driving data set.Segment cutting of driving sequence has been conducted by bayesian model-based agglomerative sequence segmentation.Based on the gaussian mixture model-latent dirichlet allocation,the lane-changing behavior primitives are extracted and their physical meaning has been analyzed.The classification model of driving styles has been set up by K-means clustering based on particle swarm optimization algorithm.Driver’s lane-changing behavior data collection platform based on real vehicles has been built to collect of drivers’ lane-changing behavior data for differentiated driving styles effectively.The generalization ability of the constructed lane-changing style classification model has been proved based on real-vehicle data.The main research content of this paper includes the following four parts:(1)The extraction of lane-changing data and the segment cutting of driving sequenceThe lane-changing behavior has been deeply analyzed from three aspects: type,process and influencing factors,and then the process of lane-changing behavior has been divided into three stages: decision,execution and adjustment in sequence of time.The screening rules of lane-changing trajctory have been formulated,and base on it,the lane-changing trajctory has been extracted from the original data set,and the trajctory has been denoised to smooth it.Then the driving sequence segment has been segmented by bayesian model-based agglomerative sequence segmentation algorithm and bayesian model-based sequence segmentation algorithm.The results have shown that bayesian model-based agglomerative sequence segmentation algorithm has better effect,so it is more suitable for this data set.(2)The extraction of lane-changing behavior primitive based on unsupervised learningGaussian mixture model-latent dirichlet allocation and multi-mode latent dirichlet allocation model have been selected to carry out unsupervised learning on the segments of driving sequence.Each segment of lane-changing behavior has been labeled,which is called lane-changing behavior primitive.Then,the fitting effect of the model has been evaluated by selected evaluation indexes.Then,the physical meaning of the primitives has been analyzed from two levels: the mapping relationship between the driver’s lane-changing behavior segments and the primitives,and the statistical feature of different lane-changing behavior primitives.(3)The construction of lane-changing style clustering modelThe driver’s lane-changing behavior has been analyzed by using the state transition probability matrix of the primitives of lane-changing behavior,and a method for dividing lanechanging stages based on primitives has been designed.A weighted driving feature considering primitive composition has been proposed and the driving feature parameters that can reflect the driver’s lane-changing style have been selected.Then factor analysis has been introduced to reduce the dimensionality of driving feature parameters.K-means clustering algorithm based on particle swarm optimization has been used to label the driving data samples,and the samples have been divided into three styles: aggressive,general,and cautious,thus realizing the characterization of the driver’s lane-changing style.(4)Verification and analysis of lane-changing style classification algorithmA data collection platform for drivers’ lane-changing behavior has been built,and drivers have been publicly recruited to participate in the data collection experiment.Then,with the driver’s lane-changing behavior data collected in the test as input,the research on the classification of driver’s lane-changing style based on the driving behavior primitive has been carried out.The generalization ability of the driver’s lane-changing style classification model has been proved.