A Fast Clustering Approach for Identifying Traffic Congestions

DBSCAN, a density-based clustering algorithm, has been widely used in pattern recognition and data mining. However, under large-scale streaming data scenarios, it suffers heavy computational cost because it examines distances between each points multiple times, especially in traffic applications which usually require calculating road network shortest distance instead of Euclidean distance. Therefore, the performance of DBSCAN for real-time clustering analyses is has become a bottleneck in such applications. Focusing on fast identifying traffic-related events, this paper utilizes linear feature to improve the efficiency of clustering by introducing linear referencing system (LRS). LRS has long been used in managing linear features, which could simplify shortest-path computation into 1-dimensional relative distance calculation, thus can significantly reduce computational complexity and cost, and meet the real-time analysis requirement of streaming data. Using vehicle GPS trajectory as an example, this study designs a LRS and its associated dynamic segmentation method for identifying traffic congestions. Experiment results proved the flexibility and efficiency of the proposed LRS-based clustering approach in identifying traffic congestions.