Assessing the Integrity of Traffic Data through Short Term State Prediction

In this study, we propose an anomaly detection algorithm on sensor traffic data. The algorithm is composed of three distinct steps: temporal detection, spatial detection, and GPS calibration. The temporal detection is based on time series analysis and detects anomalies in real-time when measured sensor values are far offset from expected readings. The spatial detector is used to prune the output of the temporal detector, identifying those anomalies which are not consistent with measurements from neighboring sensors. Both temporal and spatial prediction use the widely adopted ARIMA model. The final step is to compare the predicted speed with the average speed gathered from vehicles equipped with GPS devices and subscribed to provide their data. Experimental results on real data demonstrate that the proposed algorithm effectively differentiates between abnormal traffic events and malicious manipulation of traffic data with an average accuracy of 94%.