People Counting
We have developed a new method for people counting in video - a challenging task in computer vision. The method is general and can be applied to other objects, such as vehicles, as well.
Background & Objective: The problem of counting moving objects has many applications such as: Surveillance applications that need to count the number of people moving in a given field of view, In all cases, the objects in question, move in group which means heavy occlusion. Furthermore, most, if not all, of these systems must operate in real time, if they are to be of any use. Our goal, therefore, is to develop a general "object counting" real-time algorithm that can operate in all these scenarios, without much tuning.
Technical Discussion: Our people counting algorithm applies the mean shift algorithm to a non-Euclidean space. We start by tracking feature points in the video and compute a distance metric between these motion feature points. These feature points live in a high dimensional feature space that is not directly amenable to mean-shift clustering. Therefore, we use Multidimensional Scaling to embed the feature points in some Euclidean low-dimensional space in which we perform the mean shift clustering. The algorithm runs at about 10 frames per second.
Contact: Joseph Katz
Technology Areas:
Imaging
Computer Vision
Modification Date: September 17, 2007
