Abstract: Through the spatio-temporal anomalous flow detection technology, anomalous traffic cha- racteristics can be found in urban traffic data. Different from the single anomalous flow detection method from a time sequence, this paper proposes a k-nearest neighbor flow sequence algorithm (kNNFS) for detecting anomalous flow distributions from flow sequences. Firstly, the individual flow observation is measured in each time interval at each location. Then, a flow distribution probability database is built for each time interval at each location by calculating the frequency of the observation of a single flow. Finally, a threshold is used to determine whether the distance between a new flow distribution probability and its k nearest neighbor calculated by the KL divergence is an abnormal value. If the distance value is less than the threshold, the new flow distribution probability is updated into the historical flow distribution probability database; otherwise it is an anomalous flow distribution. Simulation results show that the kNNFS algorithm outperforms the DPMM algorithm and the SETMADA algorithm in terms of accuracy and running time.

Key words: spatio-temporal flow sequence, anomalous flow distribution detection, k-nearest neighbor, KL divergence