Research On Freeway Traffic Incident Detection Method Based-on Multi-source Information

With the continuous increase of the total mileage of the expressway,the road network structure becomes more and more complex,and the road traffic flow increases sharply.The consequent problems such as road congestion and traffic safety are increasingly prominent.Traditional freeway traffic incident detection methods are often based on a single form of traffic information,and have too many assumptions and specific model structures,which cannot meet the needs of analyzing various types of detector data.Therefore,it is necessary to explore new methods for the detection of traffic incidents so as to fully extract the traffic information contained in various types of data and improve the real-time and accuracy of incident detection.Thence,the thesis is devoted to research on freeway traffic incident detection based on a multi-source information.Firstly,a summary of different types of detection technologies is analyzed for comparing the advantages and disadvantages of various technologies.Based on the analysis,it is clear that cellphone handoff detector and microwave detectors are the research objects.The real data acquisition and data characteristic of two types of detection technologies are explored and the reliability of the two types of data is analyzed.In addition,,the effectiveness of incident detection information fusion is verified using information theory based on the principle of multi-source information fusion technology.Secondly,the characteristics of traffic flow analyzed under the incident state based on the hydrodynamic theory of traffic flow,and combining which with the data characteristics of the cellphone handoff detector and the microwave detector to establish a set of initial variables containing 12 variables.Then,a random forest-recursive feature elimination(RF-RFE)algorithm was used to determine the 8 important variables as model input parameters.The Shanghai-Nanjing Freeway field data shows that these eight parameters have obvious changes before and after the incident,which verifies the effectiveness of feature selection.Then,a PSO-LSSVM(particle swarm optimization-least squares support vector machine)incident detection algorithm is proposed based on support vector machine technology and particle swarm optimization theory.But it may be insufficient for self-adaption of single kernel function,so a PSO-MK-LSSVM(particle swarm optimization-multi kernel-least squares support vector machine)incident detection model is proposed.In order to further improve the generalization ability and detection performance of the algorithm,two different structural improvement learning models of the improved AdaBoost integrated PSO-LSSVM weak classifier were designed.In addition,a multi-class PSO-LSSVM incident type detection model proposed to detect three types of traffic incidents including traffic accidents,vehicle breakdown and bad weather.Finally,the data collected from freeway is to analyze and verify the proposed model.In the first place,SMOTE algorithm is used to solve the problem of biased data.By analyzing the evaluation indexes of DR,FAR,MTTD,CR,AUC and PI,it is proved that the PSO-LSSVM model based on multi kernel and the improved AdaBoost integrated PSO-LSSVM model with different input parameters are significant advantages in both incident detection performance and comprehensive capabilities compared with the traditional machine learning model.Among them,the improved AdaBoost integrated PSO-LSSVM model with different input parameters performs best.It also verifies the effectiveness of adding cellphone handoff detector data for incident detection performance improvement based on these two types of models.For the multi-class PSO-LSSVM incident type classification model,an ideal detection result was obtained via an example analysis.