Abstract: A Transformer based millimeter wave radar gesture recognition method was proposed for human-computer interaction of vehicle occupants. The millimeter wave radar signal was designed to be not affected by the change of light inside the vehicle, and at the same time to ensure the privacy of passengers. Firstly, the millimeter wave radar sampled signal was carried through two-dimensional Fourier transform and filtering to obtain distance-Doppler (RDM) and distance-angle maps(RAM) . Then, consecutive multi-frame RDM and RAM were fused and stitched after three-dimensional convolutional networks to obtain feature vectors. And a Transformer module was used to perform position and sequence encoding. Finally, the gesture probability distribution vector was obtained through the fully connected layer. A data set for gesture recognition classification was collected based on the actual road conditions and radar data under a variety of interference environments. The experimental results show that the method can effectively detect and recognize a variety of typical hand gestures, and the recognition accuracy can reach more than 97.14%.