Vehicle tracking has a wide variety of applications, from law enforcement to traffic planning and public safety. However, the image resolution of the videos available from most traffic camera systems, make it difficult to track vehicles, based on unique identifiers like license plates. In many cases, vehicles with similar attributes, are indistinguishable from one another, due to image quality issues. Often, network bandwidth and power constraints limit the frame rate, as well. In this paper, we discuss the challenges of performing vehicle tracking queries over video streams from ubiquitous traffic cameras. We identify the limitations of tracking vehicles individually, in such conditions, and provide a novel graph-based approach, using the identity of neighboring vehicle, to improve the performance. We evaluate our approach using streaming video feeds from live traffic cameras available on the Internet. The results show that vehicle tracking is feasible, even for low quality and low frame rate traffic cameras. Addtionally, exploitation of the attributes of neighboring vehicles significantly improves the performance.

@InProceedings\{shahri:dmsn07,
  author       = "Haidarian-Shahri, Hamid and Namata, Galileo Mark and Navlakha, Saket and Deshpande, Amol and Roussopoulos, Nick",
  title        = "A Graph-based Approach to Vehicle Tracking in  Traffic Camera Video Streams",
  booktitle    = "4th International Workshop on Data Management for Sensor Networks",
  year         = "2007",
}