Research On Urban Abnormal Event Detection Based On Taxi GPS Data

Sudden gatherings of people in cities can have serious consequences,such as serious traffic congestion or even group safety problems.To avoid such problems,the authorities prepare emergency plans in advance for large-scale events that may result in crowd gatherings,but how to quickly detect and deal with such anomalies for events that are not reported in advance or where the number of people gathered is much higher than expected is still one of the problems that need to be solved.This is a problem that has been investigated in real time,but the diversity of anomalous event types limits the efficiency and accuracy of detection.Therefore,in this paper,from the perspective of urban traffic data,we ignore the characteristics of the types of abnormal events and analyse the changes in traffic attraction in a short period of time in a region as a starting point to detect abnormal events from the attraction generated by the events to traffic,so as to achieve rapid detection and area identification of abnormal events.Firstly,this paper uses principal component analysis and community detection methods to change the data dimensions based on taxi GPS data in Xi’an,and then compares the effectiveness of different data with some neural network prediction models in taxi arrival prediction.In the experiments designed in this paper with different combinations of data dimensions and prediction models,low-dimensional data did not show better results than high-dimensional data in all time-series prediction models,and complex neural network models did not necessarily have better prediction results than relatively simple models.The combination of raw data and random forest models in this paper gave the best prediction results with a coefficient of determination R~2=0.93.Secondly,the unsupervised anomaly detection based on autoencoder is implemented using the prediction residuals.In order to evaluate the performance of the residual results of different models in the algorithm and the detection accuracy,indicators such as loss value and AUROC are introduced to evaluate the detection effect.The results show that the anomalous event detection algorithm based on autoencoder proposed in this paper can meet the usage requirements in terms of detection accuracy and efficiency.Among them,the prediction residuals of the combination of raw data and random forest model performed best in anomaly detection,and AUROC=0.86,which took less than 20 seconds to compute.Finally,anomalous occurrence areas are identified on the basis of anomaly detection results.For the detection results of different types of data,the anomaly detection method based on the community detection results and the anomaly occurrence area identification method based on the arrival change rate are proposed respectively.Based on the anomaly detection results,traffic control measures are then designed to alleviate the traffic congestion caused by anomalous events.The results show that the anomalous area location method based on community detection results is fast but has a large range of identification results;the anomalous area location method based on arrival change rate is complex but has high identification accuracy and universal applicability,which can meet the daily emergency management needs..In summary,this paper provides new ideas for urban abnormal event detection research based on taxi GPS data,providing a reliable theoretical basis and technical support for improving urban management and strengthening active early warning of urban abnormal events.