Some studies on object tracking using kalman filter and tuning of process noise covariance parameter for optimal system performance

Abstract

Object Detection and Tracking are important and challenging tasks in many computer vision applications such as surveillance, vehicle navigation, and autonomous robot navigation. Video surveillance in a dynamic environment, especially for humans and vehicles, is one of the current challenging research topics in computer vision. It is a key technology to fight against terrorism, crime, public safety and for efficient management of traffic. The present Thesis work involves designing an efficient video surveillance system in complex environments. The Kalman Filter (KF) is a widely used algorithm for tracking objects in a noisy and uncertain environment. This Thesis explores the application of the Kalman filter in object tracking and presents a comprehensive review of the theory and implementation of the Kalman filter algorithm. The Thesis also investigates different techniques for modeling the motion and measurement of objects, and examines the impact of different model parameters on tracking performance. The Thesis presents experimental results on a variety of object tracking scenarios, including tracking a single object in one and two dimensions, multiple objects, and objects with complex motion patterns. The results demonstrate the effectiveness and versatility of the Kalman filter in tracking objects in challenging environments. Finally, the Thesishighlights the importance of accurate Process Noise Covariance modeling and discusses the tuning of the Kalman filter using RMS indexwhich is verified using Sensitivity and Robustness metrics.Upon tuning of the Kalman filter using RMS index, the system performance was evaluated, showcasing a significant improvement in its overall effectiveness. In the context of a simulated one-dimensional (1D) tracking system, a comparison between an un-tuned Kalman filter and a properly tuned counterpart revealed notable enhancements. Specifically, the mean error experienced a substantial decrease from -8.51m (for an un-tuned Kalman Filter) to -0.20m (for a tuned KF), while the standard deviation reduced from 11.34m to 0.94m for the un-tuned and tuned systems, respectively. This remarkable transformation resulted in a 17-fold improvement in the system's performance for the 1D tuned tracking system. Overall, this Thesis provides a comprehensive overview of the Kalman filter algorithm and its application in object tracking, anserves as a useful reference for researchers and practitioners in the field of computer vision.