Variable-Bit-Rate Video Frame-Size Prediction by the Extended Kalman Filter Using Levenberg-Marquardt Algorithm
IEEE Transactions on Broadcasting
It is crucial to dynamically predict the future frame-sizes (bit-rates) for multimedia networking. All of the conventional bit-rate predictors are based on the assumption that instantaneous bit-rates are known precisely all the time (in the absence of uncertainty) which is surely not realistic in practice. In this work, we propose a new expectation-maximization (EM) based extended Kalman filter (EKF) to predict the bit-rates, where the EKF state-transition models will be optimized by the Levenberg-Marquardt algorithm (LMA). The main advantages of our proposed novel EKF-based bit-rate prediction approach are given as follows. First, our proposed EKF-based predictor can optimally estimate the bit-rates in the presence of uncertainty and/or noise. Second, our proposed novel EKF-based bit-rate prediction approach does not require a separate classifier to determine the individual frame-types as the conventional approach so our approach would be more robust than the conventional approach. Numerical evaluation of bit-rate (frame-size) prediction is also conducted over three movies encoded by the MPEG-4 standard. Compared to the existing Kalman-filter based bit-rate prediction methods, our proposed new LMA-EKF predictor can achieve much better performance in terms of the normalized mean square error (NMSE) and the inverse of signal-to-noise-ratio (SNR).
expectation-maximization (EM) algorithm, Extended Kalman filter (EKF), Levenberg-Marquardt algorithm (LMA), MPEG-4, multimedia network, variable-bit-rate (VBR) video streaming
Applied Data Science
Shih Yu Chang, Hsiao Chun Wu, and Kun Yan. "Variable-Bit-Rate Video Frame-Size Prediction by the Extended Kalman Filter Using Levenberg-Marquardt Algorithm" IEEE Transactions on Broadcasting (2023): 75-84. https://doi.org/10.1109/TBC.2022.3185461