Traffic Mining Applied to Police Activities

Public sector is rich of data, but this alone is not enough to fully exploit insights and information hidden in it. Data needs to be coupled with a team of data scientists and engineers, a big data platform and a legislative framework to make the famous “data driven decision making” actually possible. This is why the Digital Transformation Team introduced the Data and Analytics Framework.

Some recent advancements in the area of Mobility Data Analysis are discussed, a field in which data mining and machine learning methods are applied to infer descriptive patterns and predictive models from digital traces of (human) movement.

The research on traffic flows is historically born to improve road networks, to make trips comfortable and faster. In this research field, as in many others, literary production followed market or business demand. This paper has the objective to clarify police needs, in order to create a research request and to gain attention. It provides an organizational framework concerning needs, goals, fields and some impacts such that different areas of study can concur all together towards a pervasive comprehension of road events, a predictive Traffic Police, so that safety and security can be ensured via a targeted patrolling or intervention. The aim behind the practical level is to pave the way for the interaction between data science from one hand, and law, public administration and justice from the other hand.

Born and typically exploited in the business and industrial application domain, automatic process management, and in particular process mining, might be profitably applied also to the very different domain of traffic understanding. In facts, previous successful experiences in other movement-oriented applications have been reported in the literature. However, some peculiarities of these special domains require powerful techniques to be available. For this reason, these experience exploit the WoMan framework for workflow management, that has proved to be able to handle complex processes. This paper describes the WoMan framework along with its features and functionality, explains why it is more suitable than other process mining approaches and systems available in the current literature, and proposes a number of ways in which it might be applied to the traffic understanding domain. It also highlights possible shortcomings of the WoMan system, that might need adjustments before it can be applied at full scale on real-world traffic data.

Automatic number plate reading systems (NPRS) collect considerable amount of information from roads: number of vehicles, movements, legal status, etc. An immense quantity of information does not represent an answer to a problem if we cannot define what we are looking for and cannot extract knowledge from this information. Formal concept analysis is not recommended for big data, but it has interesting tools to extract knowledge from information stored in databases. Pruning consists in reducing initial information, done by discarding a selectable number of data that we consider not relevant. If pruned properly, the size of the database is reduced but interesting information are retained. Considerable resources are required to assess specific criminal behaviour profiles and research can help to determine which profiles we are interested in. In this paper, we focus on observed behaviour patterns in criminal activities committed in Southern Spain to reduce information provided by NPRS on Italian roads. With this reduced information we conclude that a consensus on appropriate data analysis could be reached if we focus on specific profiles.

Timely and accurate prediction of traffic flow plays an important role in improving living quality of the public, which greatly influences the polices and regulations to be enforced and abided by. In this paper, we focus on urban highway traffic prediction, and present a tensor completion based method, namely, DTC-F. It is conceived on the solid basis of dynamic tensor model for traffic prediction, and in this paper, fast low rank tensor completion and dynamic tensor structure are first combined to pursue high prediction performance. The proposed DTC-F method excavates the inner law of traffic flow data by taking account of multi-mode features, such as daily and weekly periodicity, spatial information, and temporal variations, etc. Empirical evaluation demonstrates the superiority of DTC-F, and indicates that the proposed method is potentially applicable in large and dynamic highway networks.

This research commenced when different Police departments from the Southern Spain observed an increase of irregularities in accident reports involving British vehicles. Primarily, vehicles seemed to have valid insurance, however, since they lacked Administrated Licences (Road Tax) or approved Roadworthy Certification, the damage caused in the accident by these vehicles was not being covered by Insurance Companies. British Police offers public information about British vehicles via websites for government and companies that Spanish Police checks with possible fraudulent vehicles to avoid this situation. This fraud affects countries and insurance companies. There are detected a significant amount of fraudulent vehicles that are driven in Spain. Using the information provided by British government this number of fraudulent vehicles and the problem related to accidents with non insured vehicles has been reduced. Public information and collaboration between British and Spanish Polices allow detecting frauds in vehicles with administrative irregularities and protect drivers that have accidents with these vehicles ensuring safety for road users. Methodology used in this research could be extended to different European Police departments, not only Spanish, where fraudulent vehicles are detected. This also will constitute a global reduction in risk of accidents due to mechanical problems in vehicles that are not officially checked.

This paper investigates the processing of Frequency-Modulated Continuous-Wave (FMCW) radar signals for vehicle classification. In the last years, deep learning has gained interest in several scientific fields and signal processing is not one exception. In this work we address the recognition of the vehicle category using a Convolutional Neural Network (CNN) applied to range-Doppler signatures. The developed system first transforms the 1-dimensional signal into a 3-dimensional signal that is subsequently used as input to the CNN. When using the trained model to predict the vehicle category, we obtained good performance.

The amount of traffic data collected by automatic number plate reading systems constantly incrseases. It is therefore important, for law enforcement agencies, to find convenient techniques and tools to analyze such data. In this paper we propose a scalable and fully automated procedure leveraging the Apache Accumulo technology that allows an effective importing and processing of traffic data. We discuss preliminary results obtained by using our application for the analysis of a dataset containing real traffic data provided by the Italian National Police. We believe the results described here can pave the way to further interesting research on the matter.

Traffic information plays a significant role in everyday activities. It can be used in the context of smart traffic management for detecting traffic congestions, incidents and other critical events. While there are numerous ways for drivers to find out where there is a traffic jam at a given moment, the estimation of the future traffic is not used for proactive activities such as ensuring a smoother traffic flow and to be prepared for critical situations. Therefore traffic prediction is focal both for public administrations and for the Police Force in order to do resource management, network security and to improve transportation infrastructure planning. A number of models and algorithms were applied to traffic prediction and achieved good results. Many of them require the length of past data to be predefined and static, do not take into account dynamic time lags and temporal autocorrelation. To address these issues in this paper we explore the usage of Artificial Neural Networks. We show how Long Short-Term Memory (LSTM), a particular type of Recurrent Neural Network (RNN), can overcome the above described issues. We compare LSTM with a standard Feed Forward Neural Network (FDNN), showing that the proposed model achieves higher accuracy and generalises well.