Research And Implementation Of Detection Methods Of Abnormal Events In Urban Traffic Data

With the rapid development of urban economy and the rapid advancement of urbanization,people’s travel demand is increasing and their travel modes are constantly enriched.For this reason,many large and medium-sized cities in our country are facing increasingly serious traffic problems,such as traffic congestion,accidental crowd gathering and other traffic abnormal incidents.The detection of abnormal traffic incidents is very important to intelligent transportation systems and public safety.How to accurately and real-time detect the time and place of abnormal traffic incidents has become a hot issue related to economic and social development.This issue has attracted the attention of many researchers in the field of urban computing.Compared with the traditional anomaly detection problem,the abnormal detection problem in the transportation field usually shows a complicated form in two aspects.1)Spatio-temporal complexity,that is,in order to improve the effectiveness of anomaly detection,we often need to associate various urban regions and time intervals.2)Multi-source complexity,that is,in order to improve the effectiveness of anomaly detection,we often need to comprehensively consider multiple data sources of different distributions,densities and scales.In order to cope with these challenges,this thesis designs a complete process framework for traffic anomaly detection,which is mainly divided into four phases,data preprocessing module,traffic feature extraction module,single feature anomaly scoring module,and comprehensive anomaly scoring module.In the first stage,the data preprocessing module,we partition the city map and map the original data set to each urban region.In the second stage,the traffic feature extraction module,we uses factor analysis to analyze the traffic features of adjacent spatio-temporal segments.Then we can obtain new features that are more spatiotemporal interpretable.In the third stage,the single feature anomaly scoring module,we adopt a goodness-of-fit test to score the degree of each traffic feature in each spatiotemporal segment.In the fourth stage,the comprehensive anomaly scoring module,we use the one-class support vector machine to fuse multiple data sources of adjacent spatio-temporal segments,and score the degree of anomaly of each spatio-temporal segment.Finally,this thesis conducted a large number of experiments on the real data set and artificially synthesized data set in New York City,and the extensive experiments proved the effectiveness of our anomaly detection framework.The experimental results show that the factor analysis feature extraction proposed in this thesis can effectively improve the accuracy of the anomaly detection after solving the shortcomings of original features such as information redundancy and high dimensionality.At the same time,when faced with some potential traffic anomaly detection problems,the scoring algorithm proposed in this thesis can take into account the multi-source complexity and spatio-temporal complexity,so that can effectively improve the accuracy of the anomaly detection.