Traditional aerobics gesture recognition has shortcomings such as low scalability, limitations in application scenarios, limitations in labor costs, and human–computer interaction. In order to improve the efficiency of aerobics posture recognition, based on neural network, this paper constructs aerobics posture recognition model combined with sensor network. Moreover, aiming at the unsatisfactory performance of the natural feature-based visual 3D registration method for scenes with sparse texture features and complex dynamic scenes, a deep neural network based on CNN + LSTM is proposed to establish the relative motion relationship between camera positions in continuous video sequences. In addition, this paper uses the transfer learning method to apply the network training parameters for the classification task to the video sequence posture recognition in this paper. Finally, considering the temporal correlation between video sequences, this paper uses the LSTM structure to store long-term image memory information. The experimental results show that the performance of the model constructed in this paper is good.