Traffic Anomaly Detection: Exploiting Temporal Positioning of Flow-Density Samples

It is of paramount importance to detect traffic data anomalies in a real-time manner as it helps efficient traffic control and management. Several unsupervised anomaly detection algorithms are proposed previously in the literature; however, lack of proper ground truth labels for traffic data has been always a substantial barrier to deploy and evaluate them. In this paper, we introduce a concept named Temporal Positioning of Flow-Density Samples (TP-FDS) that can be used by domain experts for fast and reliable traffic data labeling. We mathematically show that deviations in two-dimensional TP-FDS completely reflect point and subsequence anomalies previously defined in the literature of time series data. Furthermore, benefiting from this concept, we propose a novel anomaly detection framework with the help of Fast Angle Based Outlier Detection (Fast-ABOD) to be used for traffic data. Extensive data labeling experiments are conducted with the opinions of 20 different experts. Implementation of several machine learning algorithms, like KNN, OC-SVM, iForest, and LOF, is also adapted with two different setups of hyper-parameters to be used in the proposed framework. Results indicate that our framework integrated with Fast-ABOD is able to detect anomalies in traffic data better than other machine learning and state-of-the-art deep learning algorithms with more than 72% and 96% of F1 score and AUC.