A new particle filter based on interval filter and particle diffraction is proposed for the on-line estimation of non-Gaussian and nonlinear system with uncertain dynamics modeling. This algorithm computes the more accurate importance density function, which integrates the latest observations into the system state transition density, so that the approximation to the system posterior density is improved. At the same time, the workload of calculation is reduced by treating particle diffraction like light diffraction. A simulation experiment on the SINS/CNS (strap-down inertial navigation system/celestial navigation system) attitude estimation shows the effectiveness and robustness of the improved algorithm.