As virtual reality (VR) technology is widely used in various fields, particularly in dance education and training, precise dance posture capture and correction techniques are key to enhancing the quality of teaching and learning efficiency. Although current dance posture estimation technologies have made some progress, there are still deficiencies in the accurate capture of three-dimensional spatial information and real-time analysis of dynamic postures. This study addresses the limitations of existing technologies and proposes a new method combining YOLO-V3 with a cascaded convolutional neural network (CNN), aimed at improving the accuracy and real-time performance of dance motion capture in VR environments. This paper proposes an innovative three-dimensional dance pose estimation algorithm that combines YOLO-V3-based human localization with three-dimensional cropping technology and a three-stage cascaded CNN system to significantly improve the precision and accuracy of dance poses in VR environments. The method demonstrates its superiority in generating accurate human models and handling real-world data by optimizing pre-trained weights and combining synthetic and real datasets. A series of experimental verifications demonstrate that both technologies significantly improve the accuracy and real-time feedback capabilities of dance motion analysis, showcasing potential applications in the fusion of art and technology. The findings of this paper not only provide technical support for dance training in VR but also offer new research ideas and methods for intelligent analysis of complex human movements, possessing high theoretical value and practical application prospects.