Abstract: Since vehicles using fake plate violate traffic laws and regulations, infringe the rights of legal license owners and harm social benefits, it is therefore of great urgency to solve this social problem. In current detection methods, unified speed threshold is used to detect fake plate vehicles. Once an inappropriate threshold was set, it would lead to misjudgements and omissions, and thereby results in a low judgement accuracy. To solve this problem, we propose a two-phase fake plate vehicles detection framework. In the offline part, we train a distributed speed model based on sparse traffic trajectory data to calculate time-dependent thresholds for different roads. In the online part, we apply a sliding-window method to monitor whether a car is an outlier. If a car is continually detected as an outlier, it will be judged as a fake, and vice versa. We then use a real dataset to evaluate our method. The results demonstrate negative influences of noise data can be avoided by our method, so the accuracy of fake plate vehicles detection can be improved significantly.

Key words: fake plate vehicle, anomaly detection, speed distribution, sliding window