Single-particle electron cryomicroscopy (cryo-EM) involves estimating a set of parameters for each particle image and reconstructing a 3D density map; robust algorithms with accurate parameter estimation are essential for high resolution and automation. We introduce a particle-filter algorithm for cryo-EM, which provides high-dimensional parameter estimation through a posterior probability density function (PDF) of the parameters given in the model and the experimental image. The framework uses a set of random support points to represent such a PDF and assigns weighting coefficients not only among the parameters of each particle but also among different particles. We implemented the algorithm in a new program named THUNDER, which features self-adaptive parameter adjustment, tolerance to bad particles, and per-particle defocus refinement. We tested the algorithm by using cryo-EM datasets for the cyclic-nucleotide-gated (CNG) channel, the proteasome, β-galactosidase, and an influenza hemagglutinin (HA) trimer, and observed substantial improvement in resolution.