Driver’s Eye Movement Behavior Classification Algorithm Based On Improved Decision Tree And Its Application

With the rapid development of China’s road traffic industry,the number of car ownership is gradually increasing,while the number of traffic accidents is also high.In the driver-vehicle-road closed-loop system,drivers need to perceive the environment and control vehicle.Driver factors are the main factors causing accidents,traffic accidents are caused by drivers’ perception error,decision-making error or improper operation.Therefore,full observation and perception of traffic environment is the premise of decision-making and correct manipulation,which is an important way to avoid accidents.The driver can obtain 80-90% of the external information through visual perception,and accurate visual perception is the key to the stable operation of the closed-loop system.The study of drivers’ visual characteristics benefits from the driver’s eye movement data obtained by eye-tracking technology,which can reflect the changes of eye-movement behavior during driving.Then,fixation,saccade and smooth pursuit are the main eye-movement behaviors to obtain information.Therefore,this paper takes eye movement data as the main line,designs experiments to obtain drivers’ eye movement data and completes the study of drivers’ fixation characteristics and smooth tracking characteristics through the classification of eye movement behavior.The main research contents are as follows:(1)A complete framework of eye movement data analysis is proposed for eye movement related scientific research,which is suitable for any research related to eye movement data,including quality analysis of original eye movement data,eye movement data preprocessing,eye movement data classification and eye movement data post-processing.The quality of original eye movement data was analyzed before preprocessing,which can eliminate invalid data.A frequency correction method is proposed to solve the problem of irregular data frequency,which is beneficial to data analysis.(2)Considering the eye movement data acquisition depends on the hardware equipment,which leads to lack of eye movement data and the problem of noise.Aiming at the above problems,the disadvantages of general eye movement data preprocessing methods were analyzed.In the missing value filling environment,considering the fallibility of linear interpolation filling data,a Fourier series equation was proposed to fill the data.In noise reduction section,considering the traditional moving average noise reduction method in data to produce too much rush and the data is not enough for a smooth and exist white noise after noise reduction.A fusion filtering method is put forward which considers more comprehensively and can achieve better noise reduction effect.(3)For classification of eye movement data,this paper based on the movement properties of eye movement behavior.Some features for the classification of eye movement data are constructed,which take into account the movement characteristics(amplitude of motion,range of motion)and movement trend(direction of motion)of eye movement behavior.The feature set can effectively describe the differences of eye movement behavior,which is beneficial to the classification of eye movement data.Features are input to the decision tree for classification.Considering that decision tree is a prediction model,the trained model will produce over-fitting results.In order to improve this problem,this paper proposes a post-pruning method based on the tree depth and the change of accuracy after pruning which can rise to the deep decision tree.In this paper,this method improves the problem without losing the accuracy of model,make the model more simple and more efficient.The classification accuracy of fixation,saccade and smooth pursuit is 94.3%,98.2% and 71.2% respectively,which is better than other algorithms and is beneficial to the study of eye movement characteristics.(4)In this paper,drivers’ eye movement data are recorded in driving simulators and real vehicle experiments.The above methods are used to classify the driver’s eye movement data which are obtained through the experiment.Then,fixation,saccade and smooth tracking indexes are extracted from the classified eye movement data.The characteristics of driver’s preview distance are studied based on the data of fixation points,and the speed and radius of road curvature are analyzed.After the preview distance was transformed to conform to the normal distribution,the driver’s preview distance model is established by regression analysis,which was beneficial to the development of driver model.The prediction of potential collision time in drivers’ cognition is studied based on smooth pursuit point data.We design two common intersection collision scenarios.The eye movement patterns of drivers when they passed the intersection are studied and we get the eye-movement behavior transformation patterns of "saccade-fixation-smooth pursuit – chasing saccade-smooth pursuit".Then,the potential collision time is estimated by using the driver’s perspective and the smooth pursuit speed,the error between the potential collision time and the actual vehicle arrival time is analyzed.The results show that the smooth pursuit behavior is feasible to predict the collision time.