Traffic monitoring systems based on image and sequence analyses are widely employed in Intelligent Transportation Systems (ITS's) in order to analyze traffic parameters and statistics. To this purpose, tracking objects is often needed. However, occlusions can mislead a vehicle tracking system based on a single camera, thus resulting in tracking errors. In this work we present a vehicle tracking algorithm based on the KLT feature tracker which exploits a Kohonen Self Organizing Map (SOM) to drastically reduce tracking errors arising from occlusions, thus increasing the overall robustness of the system. Our method has been implemented in a real-time traffic monitoring system that has been working on daily urban traffic scenes. The experimental results we present assess the effectiveness of our approach even in the presence of quite congestioned traffic situations.