Abstract. In this document, we describe a people counting system that can precisely detect people as they are seen from a zenithal depth camera pointing at the floor. In particular, we are interested in scenarios where there are two preferred directions of motion. In our framework, we detect people using a Support Vector Machine classifier, follow their trajectory by modeling the problem of matching observations between frames as a bipartite graph, and determine the direction of their motion with a bi-directional classifier. We include experimental evidence, from four different scenarios, for each major stage of our method.