Head plane estimation improves the accuracy of pedestrian tracking in dense crowds

Human detection and tracking in high density crowds is an unsolved problem. Standard preprocessing techniques such as background modeling fail when most of the scene is in motion. Because of high levels of occlusion, dense features, and shadows, object detectors tend to produce large numbers of false detections. We introduce a new method based on 3D head plane estimation that reduces these false detections while preserving high detection rates. Our algorithm learns the head plane from observations of human heads, without any a priori extrinsic camera calibration information. In an experimental evaluation, we show that the head plane estimation technique dramatically improves the performance of a pedestrian tracker for dense crowds based on a Viola and Jones AdaBoost cascade classifier for head detection, a particle filter for tracking, and color histograms for appearance modeling.