基于VB近似的自适应δ-GLMB滤波算法

目前,基于δ-扩展标签多伯努利(δ-generalized labeled multi-Bernoulli, δ-GLMB)滤波器的多目标跟踪方法假设量测噪声协方差先验已知,而实际中量测噪声协方差可能是未知或随着环境改变而变化。针对上述问题,提出一种基于变分贝叶斯(variational Bayesian, VB)近似的自适应δ-GLMB滤波算法。该算法以δ-GLMB滤波器为基础,利用逆威沙特和高斯乘积混合分布近似量测噪声协方差和多目标状态联合后验分布,通过VB近似技术推导滤波迭代。仿真结果表明,所提算法对于线性未知量测噪声协方差场景具有很强的多目标跟踪鲁棒性,在有效估计量测噪声协方差的同时实现准确的目标数和目标状态估计。

Adaptive δ-GLMB filtering algorithm based on VB approximation

The multi-target tracking methods based on δ-generalized labeled multi-Bernoulli (δ-GLMB) filter usually assume that the measurement noise covariance is known a priori. This is unrealistic for real applications, as it may be previously unknown or its value may be time-varying as the environment changes. To solve this problem, an adaptive δ-GLMB filtering algorithm based on variational Bayesian (VB) approximation is proposed. Based on the δ-GLMB filter, the proposed algorithm approximates joint posterior density of the measurement noise covariance and multi target states by the mixture distribution of the products of inverse Wishart and Gaussian, and derives filtering iteration by the VB approximation. Simulation results indicate that the proposed algorithm has a strong robustness and could effectively estimate the measurement noise covariance and the number of targets as well as the corresponding multi target states under the unknown measurement noise covariance scenario.