Janne Heikkila and Olli Silven
Machine Vision and Media Processing Unit, Dept. EE, University of Oulu, Finland.
In this paper, we present a simple but effective tracking scheme for several visual surveillance applications such as pedestrian traffic counting and parkingarea monitoring. The tracking system developed is capable of real-time video processing in a low-power PC platform. The computational efficiency is achieved by disregarding the color information and using appropriate scaling for the images. The heart of the system is in low-level image processing, where different filtering stages are used to maintain a background model and to extract the true moving regions from spurious observations. In the upper-level, Kalman filter based trackers are used to determine the trajectories of the moving objects. These trajectories and a set of features computed from the target silhouettes are stored for post-processing. Depending on the application this data can be used to analyze the image contents or to produce figures for statistical purposes. The experiments performed with the given test material gave excellent results: all humans were detected and tracked without errors, andthe traces of the cars were broken off only when the driving direction was reversed from backwards to forwards. However, the lighting conditions in these sequences were stable which makes the problem much easier. Unfortunately, the conditions in most outdoor tracking situations are significantly worse causing the tracking problem to be often more demanding.