License plate data analytics: convoy and cohort detection

这项建议的目的是发展计算和统计方法，以便从车牌数据流中发现车队和车队。典型的车牌识别系统由摄像头组成，摄像头检测通过其视野的车牌，并记录元信息，如车牌字符、时间和车辆被看到的位置。摄像机可以是固定的（例如，安装在高速公路沿线的柱子上）或移动的（例如，安装在执法车辆上）。该项目解决了与处理来自全市或大都市地区多个摄像头的车牌读取流相关的研究问题，以协助执法人员。我们解决的两个主要问题是车队检测和队列检测。通常情况下，当一辆车被偷时，小偷会在一群人的陪同下进出犯罪现场。同样，非法毒品通常是由成群的车辆越境运输的，这些毒品是由一辆租用的或其他不知名的车辆运载的，并由犯罪分子自己的车辆陪同，而当局可能知道这些车辆。在这两种情况下，执法人员都希望确定与感兴趣的车辆“显著相关”的未知车辆。这些问题的挑战包括定义两辆车之间相关性的适当概念，以及如何将两辆或更多辆车确定为车队或队列。从这个意义上说，这些问题类似于无监督异常检测的任务，需要在给定的未标记数据集合中识别异常值。另一个挑战是数据量。典型的摄像头平均每秒记录一个车牌，而一个大都市系统可能涉及数十到数百个摄像头。必须有效地处理这些数据，以便及时向执法人员提供结果。我们的工业合作伙伴Genetec是世界领先的车牌识别系统供应商。除了直接支持研究之外，他们还提供了对来自客户部署的大型数据集的访问，用于建模和验证。

The aim of this proposal is to develop computational and statistical methods for detecting convoys and cohorts of vehicles from streams of license plate data. Typical license plate recognition systems consist of cameras which detect license plates passing through their field of view and record meta information such as the license plate characters, the time of day, and the location where the vehicle was seen. Cameras may be fixed (e.g., mounted on pillars along a highway) or mobile (e.g., mounted on law enforcement vehicles). This project addresses research problems related to processing the streams of license plate reads coming from multiple cameras over a city-wide or metropolitan region in order to assist law enforcement officials. The two main problems we address are convoy detection and cohort detection. Typically, when a car is stolen the thief is accompanied to/from the scene of the crime by a cohort. Similarly, illegal drugs are typically transported across borders by cohorts of vehicles where the drugs are carried in a rented or otherwise unknown vehicle that is accompanied by the criminal's own vehicle which may be known to the authorities. In each of these two scenarios, law enforcement officials wish to identify unknown vehicles which are "significantly correlated" with a vehicle of interest. Challenges of these problems include defining appropriate notions of correlation between two vehicles along with how correlated two or more vehicles should be to be identified as a convoy or cohort. In this sense, these problems resemble the task of unsupervised anomaly detection where one is required to identify outliers in a given collection of unlabeled data. An additional challenge is the volume of data. Typical cameras register, on average, one license plate per second, and a metropolitan system may involve anywhere from tens to hundreds of cameras. This data must be handled efficiently to provide results to law enforcement officials in a timely manner. Our industrial partner, Genetec, is a world-leading provider of license plate recognition systems. In addition to direct support of research, they are providing access to a large dataset from one of their customer deployments to be used for modeling and validation.