A video surveillance system based on interacting multiple models

The video-based surveillance systems are becoming widespread due to the increasing security needs. Consequently, these systems bring huge volumes of visual data to be analyzed. The automated systems are developed to assist human operators in time-consuming scene analysis. Besides, they enhance the surveillance efficiency by tracking interesting moving objects and interpreting the tracking results for potentially dangerous situations or suspicious activities.In this thesis, we present an automated visual surveillance system with real-time and robust tracking capabilities. The system detects moving objects under changing background conditions by an adaptive Mixture of Gaussians method. The detected objects in the consecutive video frames are properly associated with each other by means of data validation and association algorithms. The tracking algorithm makes use of the prediction and estimation results of the Interactive Multiple Modal (IMM) estimator operating on constant velocity and coordinated turn motion models simultaneously. The system has been used to analyze PETS 2001 datasets which provide a unique test environment for the objective evaluation of the tracking algorithms.