Performance comparison of Kalman Filter and extended kalman filter for human tracking and prediction with particle swarm optimisation

Abstract

In the past two decades, object tracking has progressively advanced in computer vision and image processing. Tracking is a collection of algorithms that detect and track objects in a video sequence. This has resulted in a broad variety of applications, including surveillance, biometric identification or biological imaging, human-machine interactions, traffic control, and intelligent vehicles, all of which have benefited from tracking applications. This paper presents an IoT-based human tracking system that employs the Kalman filter (KF) and extended Kalman filter (EKF) algorithms. However, the performance of the filters is impacted by noise. Therefore, instead of manually tuning the KF and EKF’s process noise and measurement error, particle swarm optimisation (PSO) is utilised to optimise them. Four 2-dimensional models (namely conventional KF and EKF, optimised KF-PSO, and EKF-PSO) were developed and evaluated using ten distinct human sets of data containing 100 samples each. The object’s tracked positions are estimated in horizontal and vertical directions. Accuracy analysis was used to compare the four models’ quality performance. With an average mean square error of 5.99 mm (or 0.599%), the EKF-PSO model outperformed the conventional EKF model, which had an error of 7.18 mm (or 0.718%). A 17.2 mm (or 1.72%) error was found in the KF-PSO model. The last one is the traditional KF model, yielding an error of 22.4 mm (or 2.24%). As a result of its higher accuracy, the EKF-PSO model outperforms the other three models.