Research On The Traffic Flow Data Based Method For The Detection Of Abnormal Traffic Event

Traffic anomaly detection is an important part of the traffic safety supervision and control system,and also a hot topic in the field of transportation research.However,the real-time complexity and contingency of road traffic have brought a great challenge to the detection of abnormal traffic events.The occurrence of abnormal traffic events will lead to apparent disturbance of traffic flow parameters such as vehicle speed,occupancy,and traffic volume,and this phenomenon can be used for the automatic detection of traffic anomaly.Based on the traffic flow data and events records collected from PeMS system of the I80-E freeway,we studied the automatic detection of traffic anomaly based on traffic flow parameters.The main contribution is listed as follows:(1)Referring to the existing algorithms and the basic understanding of the traffic system,we proposed a series of statistical-based or wavelet-based methods to extract the information about traffic congestion,abnormal traffic flow,and state instability adaptively.A comprehensive characterization of real-time traffic state is provided by the extracted information.Meanwhile,the magnitude difference of parameters among different detectors is eliminated through these processes,which makes it possible to compare the states of different detectors at the same level.(2)Based on the samples collected from the historical data of I80-E freeway during the year 2016.For training of the vehicle collision detection model,six different algorithms including linear discriminant analysis,quadratic discriminant analysis,neural network,support vector machine,decision tree and random forest are used.The final model is constructed after the process of feature extraction,feature selection,parameter optimization,and model selection.This model is then applied to the traffic data of the I80-E highway during December and February 2017 for validation.The correctness of our model is proved by the successful detection of 46 cases in 58 vehicle collision events in the new data.Through the experimental research and practice in this thesis,it is proved that the traffic flow state(such as road congestion,abnormal flow,and state instability)obtained by the proposed methods,can assist the machine learning of abnormal traffic events detection effectively.We believe the outcome of our study will provide a lot of inspiration and support,with significant scientific research value and social meaning,for further work by other researchers in the relevant domain.