Abstract: ‍ ‍One of the most popular research fields of millimeter wave radar is to recognize human behavior by means of convolutional neural networks in deep learning. Due to the defect of convolutional neural networks’ structure， each layer of convolutional neural networks can only extract the relationship between the internal pixels of the receptive field， and lacks the ability to extract global features. Moreover， the number of public radar data samples used for human behavior recognition is too small at present， and the recognition accuracy of traditional depth learning algorithms for radar micro-Doppler images is not high， and the phenomenon of over fitting is easy to occur in the training process. To address these issues， a radar-based human behavior recognition algorithm based on Snapshot Ensembles and Transfer Learning， is proposed in this paper. First of all， in order to solve the problem that the traditional convolutional neural networks cannot extract the global features of images in the field of deep learning， this algorithm introduced attention mechanism by building a model named Vision Transformer （VIT）. After that， the public natural image dataset named ImageNet， was used to perform task migration and feature space migration on the Vision Transformer model， which aimed to solve the problem of identification overfitting of radar micro-Doppler images. Last but not least， the voting mechanism algorithm based on Snapshot Ensembles was used in this paper to improve the recognition ability of the Vision Transformer model， though radar micro-Doppler images have complex background information. In a word， the experimental results show that， in case of too few samples capacity and complex background of target task data set， this algorithm can effectively improve the recognition accuracy of radar micro-Doppler images without increasing the training cost. Using this algorithm on Vision Transformer model， the recognition accuracy rate of the algorithm for radar micro-Doppler reached more than 89.25%， which is better than the typical convolutional neural networks.