Speaker
Description
This paper presents an adaptive Kalman filter for video-based object tracking, in which the covariance matrices of the process and measurement noise are dynamically updated in real time. The proposed approach is based on a linear stochastic model without introducing a control input, which is justified by the nature of video observation, where only measurement information from consecutive frames is available.
The state vector includes object coordinates, velocities, and accelerations, enabling a more accurate representation of the system dynamics and improved prediction under varying motion regimes. The adaptation of the covariance matrices is achieved through innovation analysis and the application of a statistical consistency criterion (NIS), providing automatic balancing between the model and measurement contributions.
The effectiveness of the proposed method is evaluated through simulation studies under different levels of measurement noise. The results demonstrate a significant reduction in the root mean square error of the estimates and robust filter performance under varying measurement conditions, including noise and temporary loss of observations