Video Analytics Using Detection on Sparse Frames

The paper considers two problems of video analytics, which can be solved by tracking people in a video stream: people counting and estimation of queue waiting time. Modern video surveillance systems have several hundred thousand cameras, which is why one of the most important problems that video analytics has to face is the optimization of computing resource usage. Most presently available tracking algorithms are inefficient because they use computationally expensive CNN-based detectors on frequent video frames. In this paper, we propose methods for solving the problems mentioned above, which improve overall efficiency by applying detection on sparse frames. The experimental evaluation of the proposed methods shows their consistency in terms of both performance and computing resource usage.