The Terrascope Dataset: Scripted Multi-Camera Indoor Video Surveillance with Ground-truth

This paper introduces a new video surveillance dataset that was captured by a network of synchronized cameras placed throughout an indoor setting and augmented with groundtruth data. The dataset includes ten minutes of individuals who are moving throughout the sensor network. In addition, three scripted scenarios that contain behaviors exhibtied over a wide-area, such as "gathering for a meeting" or "stealing an object" are included to assist researchers who are interested in wide-area surveillance and behavior recognition. In addition to the video data, a face and gait database for all twelve individuals observed by the network of cameras is supplied. Hand-segmented ground-truth foreground regions are provided for every 500th frame in all cameras and for many sequential frames in two overlapping views. The entrance and exit time of each individual in each camera for one of the scenarios is provided in an XML database. We believe that the dataset will help provide a common development and verifcation framework for the increasing number of research efforts related to video surveillance in multiple, potentially non-overlapping, camera networks.