Multi-dimensional convolution resident multi-head attention action recognition method based on multiple sensors

The invention discloses a multi-dimensional convolution resident multi-head attention action recognition method based on multiple sensors. The method comprises the steps of sensor signal acquisition and preprocessing, key feature extraction, multi-dimensional convolution resident multi-head attention network construction, multi-dimensional convolution resident multi-head attention network training and multi-dimensional convolution resident multi-head attention network testing. According to the method, multi-sensor binary anonymous data is adopted, so that the problems of large video data volume and privacy disclosure are solved; a multi-dimensional convolution module is adopted, so that the problems of large feature data volume and low feature extraction precision in the field of action recognition are solved; a time window segmentation module is adopted, and a multi-head attention classification module is resided, so that the problem of low sensor channel information fusion degree is solved, the action recog