A Novel Spatio-Temporal Approach to Handle Occlusions in Vehicle Tracking

In Intelligent Transportation Systems (ITS’s), the process of vehicle tracking is often needed to permit higher-level analysis. However, during occlusions features could “jump” from an object to another one, thus resulting in tracking errors. Our method exploits second order statistics to assess the correct membership of features to respective objects, thus reducing false alarms due to splitting. As a consequence, object’s properties like area and centroid can be extracted stemming from feature points with a higher precision.We firstly validated our method on toy sequences built ad hoc to produce strong occlusions artificially and subsequently on sequences taken from a traffic monitoring system. The experimental results we present prove the effectiveness of our approach even in the presence of strong occlusions. At present, the algorithm is working in the daytime in the Traffic Monitoring System of the city where we have been living.