Research On Driving Scene Complexity Based On Supervised Learning

With the continuous development of road traffic systems and the acceleration of urbanization,driving scenarios are becoming more and more complex,and the driving environment contains a large amount of traffic information,composed of many static and dynamic factors,and the uncertainty is extremely high.The driver’s accurate cognitive assessment of the complexity of the surrounding scene can reduce the occurrence of accidents,and in the complex road characteristics and external environment,the driver’s information load needs to perceive and judge increases,which is not conducive to safe driving.Therefore,studying the complexity of driving scenarios and analyzing the correlation between scene element composition and scene complexity are of great significance for improving traffic safety and driving experience,and can also provide a theoretical research basis for the timing and decision-making of man-machine co-driving control switching.Based on Project Cars,this paper builds a scene library covering general scenes and special extreme scenes,designs driving experiments,and builds an experimental environment to obtain research data.Based on computer vision and other methods,the characteristics of scene elements are obtained,the complexity level is marked with discrete numerical labels,and the complexity prediction model of driving scene is built based on supervised learning,which reveals the mapping relationship between the influencing factors of driving scene complexity and the complexity level.First,a brief overview of the relevant ideas and implementation methods of the research content in this paper is given.Based on the existing research on scene complexity,the complexity of driving scenarios is defined in detail.After classifying and sorting out the various factors affecting the complexity of driving scenes,9 game tracks were selected,and the scene element library,including road alignment,environmental elements and traffic flow state,was constructed by combining different scene elements,and driving simulation experiments were carried out.The image recognition method and marking method are used to obtain the element parameters such as road length,width,turning radius,weather lighting,etc.,and the collection of driving scene characteristic parameters is generated.Secondly,relying on manual labeling and image recognition technology,driving event data and vehicle status data,including driving events and speed data,are obtained.Considering the indicators such as collision degree,sharp braking behavior and vehicle rushing off the track,the complexity of different driving scenarios is fully reflected,and an index system suitable for characterizing the complexity of driving scenarios is constructed.After calculating the index weight by entropy weight method,the complexity score of each driving scene is obtained to quantify the complexity of each scene.Based on the K-means algorithm,the driving scene is divided into three levels,and the complexity of the driving scene is labeled to more intuitively represent the complexity of the driving scene.Finally,the driving scene complexity prediction model was constructed by using SVM model,decision tree model,Bayesian neural network and linear classifier using SVM model,decision tree model,Bayesian model,BP neural network and linear classifier in supervised learning,in which the SVM prediction accuracy reached95.2%,the decision tree prediction accuracy reached 93.3%,and the Bayesian prediction accuracy reached 91.4%.The prediction accuracy of the linear classifier reached 83.8%,and the prediction accuracy of the BP neural network reached 94.8%.