To address the problems in visual navigation scenes, like dependence on scene texture features and low accuracy and robustness due to moving objects, we propose a two-stage coarse-to-fine position estimation method based on the neural radial field (NeRF), which realizes a visual high-precision position estimation method based on the implicit representation of neural networks. Specifically, the initial position estimation network Rnet is constructed based on the channel attention mechanism, which is used to estimate the initial position under the condition of large misalignment angle; then a depth correction module (DCM) is designed based on the constraint of trusted zone and the interest point sampling strategy, which provides more accurate priori information for the precise estimation of the position by optimizing the depth of the NeRF model; finally, a PoseNeRF network is constructed using gradient descent based on SE (3) manifold to minimize the residuals between NeRF pixel color, depth inference values, and true values, achieving accurate pose estimation. In the UE simulation dataset experiments, the position accuracy is better than 0.0057 m, and the attitude accuracy is better than 0.0190°; in the Kitti and OVIT dataset experiments, the position accuracy is better than 0.0598 m, and the attitude accuracy is better than 0.0373°. Comparatively, our algorithm shows nearly a six-fold improvement in the convergence speed of bit-posture estimation when benchmarked against other algorithms such as iNeRF and NeRF--, showcasing its superior robustness and efficiency in visual navigation tasks.