In this paper, a vision-aided terrain referenced navigation (VATRN) algorithm constructed by point-mass filter (PMF) is accelerated by graphic processing unit (GPU). The terrain referenced navigation algorithm estimates the vehicle's position by blending INS data with measured terrain height, and matching that data with the stored digital terrain elevation database (DTED). On the other hands, the VATRN algorithm obtains odometry data from visual sensors instead of inertial sensors. The odometry data is estimated by the homography relationship of two successive ground images of a monocular camera. Point-mass filter is one of the TRN algorithm based on the Bayesian estimation theory, and it contains convolutional integral of each points for the time update process. The convolution is the computational burden and can be accelerated by parallel computing to improve the estimation performance of PMF with sufficient grid points. GPU is employed to accelerate the PMF and numerical simulations are performed to analyze and evaluate the performance of the proposed method. The results show that the precise autonomous navigation of unmanned aircraft is achieved by the accelerated vision-based TRN algorithm.