A Comparison of Methods to Detect People Flow Using Video Processing

We study a set of methods to detect people flow using video processing. As a source, we use surveillance cameras, located above pedestrian zones. As a basic approach, we have chosen detection of individuals with tracking algorithm, based on Kalman filtering. For the study, we have chosen the following detectors: ACF (Caltech), ACF (INRIA), Viola-Jones, and Histogram of Oriented Gradients (HOG). We compared the results of the detectors with a manual counting of people in the frame. The numerical experiments have shown that the accuracy of calculations depends on the direction of the flow, crowd density, and frame size. For tested video fragments, ACF algorithms have shown the best results. We also performed a statistical analysis of detecting errors.