Abstract: In order to address the issue that the performance of tracking using the traditional Poisson Multi - Bernoulli Mixture(PMBM) filter rapidly deteriorates when the measurement noise covariance is unknown, an improved extended target tracking algorithm and its implementation method was proposed. The algorithm constructs a Poisson multi-Bernoulli mixture filter model based on the variational Bayesian(VB) framework,which assumes the measurements of an extended target are generated by the measurement generation points on the surface of the target. It jointly estimates the posterior probability density of the measurement generation points and the measurement noise covariance. The variational Bayesian approach is implemented to approximate the joint posterior probability density as a Gaussian inverse gamma distribution. During the filtering process, the parameters of the Gaussian inverse gamma distribution are iteratively updated to achieve the goal of jointly estimating the states of the measurement generation points and the measurement noise covariance. Finally, the state of the extended target is obtained by clustering the measurement generation points. Simulation experiments show the proposed VB-PMBM algorithm in this paper can effectively estimate the states of multiple extended targets, and its tracking performance is superior to that of the traditional PMBM algorithm.