Fast single-view people tracking

Abstract

This paper presents a fast people tracking technique comprising a simple background subtraction and Gaussian mixture PHD filtering on a constrained 3D motion model. Our technique is well-suited for video surveillance applications where behavior analysis and event detection are required without having to identify and track each individual. Comparisons with two state-of-the-art visual tracking methods and BraMBLe, a well-known recent technique, show that our method is substantially faster while exhibiting generally better tracking performance. The speed improvement is achieved by the use of simple background subtraction which saves computational resources and the exploitation of temporal information via PHD filtering, which compensates for the information loss incurred in the background subtraction at a fraction of thecost of a good detection scheme.