Abstract—Understanding the urban traffic flow at intersections is helpful to formulate traffic control strategies, so as to ease traffic pressure and improve people's living standards. There are many related researches on traffic flow, and similarity research is one of them. Different from the traditional way, this paper studies the traffic flow from the perspective of image similarity. The Convolutional Variational Auto-Encoder (CVAE) is introduced to extract the low-dimensional features of traffic flow during a day, and Affinity Propagation (AP) clustering algorithm is used to cluster the features without real labels. Combining the clustering results with geographic coordinates reveals the distribution pattern of traffic flow. The experimental data includes about 10 million vehicle records at 650 intersections in Changsha on a certain day. The clustering results show that the traffic flow at the intersection of Changsha City can be divided into three categories according to the time-variant trends, and the distribution of each category basically conforms to the daily traffic laws of the city. Furthermore, the effectiveness of the clustering process is further verified by clustering the open source temporal data of different lengths.

Index Terms—traffic flow, convolutional variational auto-encoder, feature extraction, affinity propagation clustering algorithm

Cite: Min-Tong Su, Jin Zheng, and Zu-Ping Zhang, "Clustering Mining of Urban Traffic Flow Based on CVAE" Journal of Traffic and Logistics Engineering, Vol. 8, No. 2, pp. 34-44, December 2020. doi: 10.18178/jtle.8.2.34-44

Copyright © 2020 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC BY-NC-ND 4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.