**1. Introduction**

With economic development and the continuous expansion of the scale of cities, the number of vehicles increases sharply, inducing a frequent occurrence of road traffic offenses. In the urban traffic system, the supervision of abnormal vehicles, such as fake license plates, suspected of cases of illegal operation and other anomalies, have been viewed with high precaution by the traffic administration since they seriously threaten normal traffic order and safety. Intelligence analysis about these abnormal vehicles makes broader sense for assisting effective vehicle monitoring and management. The intelligence information includes all the holographic states of a vehicle, for example, the basic data of the vehicle and driver, the trajectory, the origin destination (OD) characteristics, aggregation with other vehicles, and others. It is really an important research topic on the real-time intelligence information analysis method by current devices deployed onboard or roadside.

In many intelligence information types mentioned above, the trajectory of the vehicle contains abundant spatial and temporal distribution features. Using a geographic information system (GIS), the trajectory can be accurately projected on an electronic map to give an intuitive presentation of the vehicle's movement. Moreover, through the analysis of massive trajectory data, the trip rules of vehicles, such as the OD, and the travel preferences of the driver can also be obtained. These are very valuable for intelligent macro traffic management. In some microscopic traffic control applications, the vehicle trajectory is also meaningful. For example, the single trajectory prediction can be used to provide support for capturing illegal vehicles [1,2], while multiple trajectories prediction can be used to evaluate the short-time traffic volume for designing a proper traffic signal plan [3–8].

Using the electronic surveillance cameras deployed on roads, the driving status of a vehicle such as the time, license plate number, speed, lane and direction can all be acquired. On the basis of the road network topology, a vehicle passes through a series of intersections and the trajectory of the vehicle can be obtained easily from the driving status data collected by a video-imaging detector at the intersections. Compared with the methods based on location-based-services (LBS) [9–12], wireless fidelity (WiFi) probes [13–15] and cellular signaling [16–19], vehicle trajectory extracted form vehicle license plate data has certain advantages in terms of wide adaptability, accuracy and visualization effects. Fully using the driving status acquired by the video-imaging detectors, this paper studies the vehicle trajectory prediction scheme based on the latent travelling rules extracted from massive vehicle license plate data. For a certain vehicle, the turning characteristic at an intersection is described by a turning probability transition matrix in which the probability is calculated according to statistics of historical trip chains from the vehicle license plate. The experimental study shows that the proposed method presents a better performance in a short-term trajectory prediction.

The rest of the paper is organized as follows. Section 2 presents related work. In Section 3, trip chain building and compensating methods based on vehicle license plates are presented. In Section 4, the turning state transition matrix based on historical trip chains is proposed for trajectory prediction and the one/*k*-step trajectory prediction models are described. The experimental study and the evaluation results are presented in Section 5. In Section 5, we conclude the paper and provide directions for future work.

## **2. Related Work**

Vehicle trajectory is important intelligence information for both urban macroscopic traffic managemen<sup>t</sup> and microscopic traffic control. Generally, vehicle trajectory uses a vehicle's or driver's location and identity as major data foundations. According to different vehicle or driver locations and identity acquisition types in an urban traffic environment, the current trajectory-building methods can be classified into the following categories: LBS-based methods, WiFi probe-based methods, cellular signaling-based methods and video-imaging-based methods.

The LBS-based method mainly uses global positioning system (GPS) floating vehicle data to track the target vehicle. The movement of the vehicle is detected continuously in time and the trajectory can be presented visually combined with the application of an electronic map. In [9], based on the vehicles' historical trips, the relationship between different road segments are built by transforming the road network model into a matrix, and the driving regularity of vehicles is analyzed for the design of the algorithm of vehicle route prediction. In [10], the historical vehicle GPS data is used to match the current trajectories and infer future possible destinations. The method predicts the trajectory by a systematic procedure for describing the features of the similarity of trajectories and destinations. The method also takes the station correlations and the user historical destinations into account. The positive prediction rate of the proposed method can reach 92% under the condition that the test trip has been completed over 70%. In [11], authors propose a vehicle trajectory prediction method based on the hidden Markov model. The relevant parameters are analyzed by historical vehicle GPS data, and the Viterbi algorithm is used to seek the double layers hidden states sequences corresponding to the recent driven trajectory. The future vehicle trajectory is predicted by a novel algorithm based on the hidden Markov model of double-layer hidden states. In [12], using the GPS data, a vehicle trajectory prediction method based on a variable-order Markov model is proposed. Kernel smoothing which combines sequence analysis with the Markov statistic is used for model building. The method presents a higher performance in prediction accuracy. However, these methods are only suitable for some special commercial vehicles

which are equipped with GPS or other onboard positioning devices, and not applicable for most private vehicles.

By installing certain WiFi probe devices [13] at intersections or roads, the WiFi probe-based method generates the trajectory by detecting the passing time and the media access control (MAC) addresses of electronic terminals with WiFi-connecting function, such as the onboard unit and driver's mobile phone. In [14], authors use a WiFi probe as the data collector to scan the mobile devices within a certain range in a certain period of time to obtain the MAC address, reference distance, time stamp and other information of mobile phone. Furthermore, a customer flow prediction model based on seasonal auto regressive integrated moving average (SARIMA) model and BP neural network model is built. In [15], an urban mobility trajectory analysis model based on large-scale WiFi probe request data is built. Unique entries per access point and per hour of WiFi data are aggregated to approximate local population counts by type of user. In the model, spatial network analysis is used to apply the results to the road and pedestrian sidewalk network to identify usage intensity levels and trajectories for individual street segments. The research demonstrates the significant potential in the use of WiFi probe request data for understanding mobility patterns. Similar in [16], authors design a user feature space in which frequent trajectory patterns are used to represent each user as a feature vector based on the anonymized WiFi scan lists. As per the popularity of electronic terminals, the application of WiFi probe is more adaptable than the LBS-based method. However, the identity of the target is denoted by the MAC addresses of the electronic terminal. It is not directly relevant to the vehicle itself. Besides, in order to achieve an urban-wide trajectory analysis, several WiFi probe devices should be built, inducing high expense for construction.

As with the WiFi probe scheme, the cellular signaling-based method also uses the location and identity of electronic terminals to generate the trajectory. By contrast with the former, the location is detected by the mobile station and the identity is generally the mobile user or the MAC of the mobile terminal. For example, in [17], the authors propose a data-driven method for dynamic people-flow prediction based on cellular probe data. The grid-based data transformation and data integration module is proposed to integrate multiple data sources for human daily trajectory generation. Moreover, a dynamic people-flow prediction model based on random forest is also presented. The experimental results show that the proposed method can provide prediction precision of 76.8% and 70% for outbound and inbound people, which is better than the single-feature model. In [18], the authors introduce a mobility modeling method based on real tra ffic data collected from 4G cellular networks, including data collection, trajectory construction, data noise removal, data storage and analysis. The experiments discover the user's mobility features, changing of city hotspots, and mobility patterns. However, locating using a mobile cellular network and mobile base station is inaccurate outdoors without supplementary GPS or WiFi devices. Hence, the precision of these methods is relatively low and they are only suitable for microscopic tra ffic and population evolution analysis.

Fully using the driving status data collected by electronic surveillance cameras on roads, the vehicle trajectory is acquired by time detecting and license plate number series. In [19], an o ffline method for historical OD pattern estimation based on automatic license plate recognition data is proposed. A particle filter is used to estimate the probability of a vehicle trajectory from all possible candidate trajectories combined with the time geography theory. In this method, the path flow estimation process is conducted through dividing the reconstructed complete trajectories of all detected vehicles into multiple trips. The proposed method is verified and the results show that the MAPEs of the OD estimation are lower than 19%. In [20], a vehicle trajectory extraction algorithm based on license plate recognition data is proposed. The license plate and timestamps are used for the establishment of trip chain. Aiming at the data loss problem when detecting a vehicle license plate, the K shortest path algorithm and gray relational analysis are further used for trip chain compensation. The research focuses on extracting the vehicle trajectory and the prediction of future driving state is not studied. In [21], authors propose a vehicle trajectory reconstruction method based on license plate data. In the method, travel time threshold is used to obtain a single travel chain and the similarity of the ideal

solution and depth first search method are used to build a vehicle trajectory reconstruction model. It can effectively solve the problem of incomplete license plate number data. However, the related research mainly focuses on the macroscopic trajectory modeling and OD analysis and is seldom concerned with the microscopic real-time vehicle trajectory prediction.
