Abstract
This paper presents an innovative approach to vehicle tracking in video streams, utilizing graph structures, integrated prior knowledge, and forecasting techniques. The method encompasses statistical analysis and prediction of non-identical periodic events, specifically tailored to the characteristics of vehicles in motion. Through the generation of pseudo-observations, auxiliary information is provided to the tracker, effectively capturing the underlying dynamics of vehicle movements. The approach is evaluated on tracking datasets, demonstrating substantial improvements in tracking performance, especially in scenarios with camera sensor occlusion caused by physical obstructions, which is a common challenge encountered in camera-based tracking systems.