Computationally Efficient CPHD Fusion based on Generalized Covariance Intersection

This paper considers distributed multi-target tracking based on the Cardinalized Probability Hypothesis Density (CPHD) filter and the Generalized Covariance Intersection (GCI) fusion rule. For a distributed sensor network which has limited processing power and computational capability, the Gaussian-mixture-based CPHD (GM-CPHD) fusion is more parsimonious and practical compared to Monte Carlo-based CPHD fusion. Hence, the GM implementation is considered in this paper. Nevertheless, the GM-CPHD fusion is still characterized by a computational complexity that grows exponentially with the number of sensors, and high-order polynomially with the number of targets, This work focuses on devising a computationally efficient GM-CPHD fusion algorithm so as to enhance the practical applicability of CPHD fusion. To this end, the fused CPHD is approximated as a weighted sum of fused CPHDs, each obtained by performing fusion with respect to a smaller group of components. Based on the proposed approximation, we further devise a parallelizable CPHD fusion that can reduce the computational complexity of the original CPHD fusion, at the price of a slight performance loss. Further, by exploiting the Union-Find data structure, an efficient grouping procedure that can be performed at an early stage is proposed. The performance of the proposed method is demonstrated via simulation experiments in a challenging scenario.