Pylon grid: A fast method for human head detection in range images

Abstract

We present a fast and accurate method for human head detection in range images captured by a stereo camera that is positioned vertically, pointing from the roof to the ground. We show how a static grid of measure points (pylons) can outperform hill climbing techniques and how a fast median filter can be used for effective preprocessing of the range data. The Pylon Grid algorithm detects all local minima in the range image and has a linear time complexity with respect to the number of pylons. One important prerequisite for applying the Pylon Grid algorithm to human head detection is a one-to-one relationship between human heads in the scene and local minima in the range image. This is achieved in a preprocessing phase, where an orthographic projection, convexization and noise filtering is applied to the range data. The preprocessing steps also run in linear time and can be parametrized for a further trade-off between computational cost and accuracy. The method was tested with crowded scenes, where multiple dense groups of up to six people move in random directions and have physical contact.