How much bandwidth does surveillance system require?

One of the main challenges in surveillance systems lies in the massive amount of video involved in providing potential key content with sufficient resolution. This paper shows that there exists a sweet spot, which we term critical video quality that can be used to reduce bitrate of video transmission without significantly affecting the accuracy of the surveillance tasks. We present a new city surveillance dataset which was divided into three types of scenarios, and we analyze subjective data collected via human subjective testing for object identification. These data are then used to create objective measurements (models) to drive video compression ratio based on the detection probability. The main idea is to find out the lowest bitrate of video transmission while maximizes the probability of detecting objects which are carried or abandoned. Experiment results shown that our generalized models can predict acceptable video quality for object identification in rational ways.