Abstract: NeRF(neural radiance fields) is a neural network-based 3D reconstruction technology. It defines a scene as a five-dimensional radiance field function of position and viewing angle, represented through an implicit neural network. This technique only requires two-dimensional images of a single scene from different perspectives. Coupled with volume rendering equations, a neural radiance field model of the scene can be trained through deep learning. This model can be used to synthesize high-quality images from new perspectives. This paper surveys and categorizes existing work on NeRF, mainly introducing the basic principles and advantages of various methods from aspects such as relative pose estimation, view aggregation standards, and rendering process optimization. The focus is to elucidate the similarities and differences between different methods to help understand their relationship. The paper also discusses various application scenarios that benefit from the NeRF synthesis mechanism. Through the analysis of the advantages and limitations of existing methods, the paper provides a perspective on future research trends.