STTR

Abstract

To fully exploit the capabilities of sensors in real life, especially cameras, smart camera surveillance requires the cooperation from both domain experts in computer vision and systems. Existing alert-based smart surveillance is only capable of tracking a limited number of suspicious objects, while in most real-life applications, we often do not know the perpetrator ahead of time for tracking their activities in advance. In this work, we propose a radically different approach to smart surveillance for vehicle tracking. Specifically, we explore a smart camera surveillance system aimed at tracking all vehicles in real time. The insight is not to store the raw videos, but to store the space-time trajectories of the vehicles. Since vehicle tracking is a continuous and geo-distributed task, we assume a geo-distributed Fog computing infrastructure as the execution platform for our system. To bound the storage space for storing the trajectories on each Fog node (serving the computational needs of a camera), we focus on the activities of vehicles in the vicinity of a given camera in a specific geographic region instead of the time dimension, and the fact that every vehicle has a "finite" lifetime. To bound the computational and network communication requirements for detection, re-identification, and inter-node communication, we propose novel techniques, namely, forward and backward propagation that reduces the latency for the operations and the communication overhead. STTR is a system for smart surveillance that we have built embodying these ideas. For evaluation, we develop a toolkit upon SUMO to emulate camera detections from traffic flow and adopt MaxiNet to emulate the fog computing infrastructure on Microsoft Azure.