Space-Time Vehicle Tracking at the Edge of the Network

While large number of cameras have been deployed on campus, governments and restricted areas to protect the safety of residents and their properties, manually searching for the track of a suspicious human or vehicle is painful and prone to errors. An intelligent camera surveillance system is crucial to make the usage of the large-scale camera system efficient and smooth. In this paper, we present our Space-Time Vehicle Tracking system (STVT) which targets to track all vehicles over time and store the result as trajectories of vehicles in a graph database. We are then able to respond to queries such as Where did this suspicious vehicle come from? and Where is it headed to? directly from the stored trajectories. STVT makes use of computer vision techniques including Yolo v2, Dlib's correlation tracker, and adaptive histogram to transform each video stream into a stream of vehicle detection events. A publish-subscribe-based messaging system is built above the camera networks to enable communicating vehicle detection events between cameras. A graph database is used to store the vehicle detections and connect them into trajectories of vehicles. We have a built a prototype of STVT and are currently assessing its effectiveness with a small subset of cameras. Specifically, we process video streams from 7 Georgia Tech's on-campus street cameras with STVT, and identify new challenges and future works in a real-world deployment.