View synthesis is a fundamental task in computer vision, known for its significantly higher complexity compared to conventional vision problems. The introduction of Neural Radiance Fields (NeRF) marked a major breakthrough in this field, substantially improving previous methods and pushing view synthesis to unprecedented levels. This survey aims at systematically reviewing the progress of NeRF-based models in computer vision. We begin by explaining the core principles underlying the success of NeRF. Then, we delve into and analyze seven representative NeRF-based representation forms, including Implicit Representation, Neural Point Cloud, and others. Next, we provide a comprehensive comparison and analysis of 14 major research directions that enhance NeRF, such as Modeling Different Practical Capturing Scenarios, Generalization in Modeling, and Modeling Dynamic Scenes. In addition, we conduct both qualitative and quantitative evaluations of numerous NeRF-based methods on multiple datasets, comparing training time, rendering speed, and memory requirements. Finally, we discuss potential future research directions and challenges in this field. We hope that this work will inspire further interest and contribute to advancing the application and development of NeRF in computer vision.