The invention discloses a representation method of human motion, which maps the human motion to a low-dimensional embedding space through nonlinear dimension reduction; models the data after dimension reduction with a linear time series model. Among them, the human motion is mapped to a low-dimensional embedded attitude space through nonlinear dimensionality reduction; the realization is: z=(x1+jy1, x2+jy2,...,xM+jyM)T. Model the dimensionally reduced data with a linear time series model; implementation steps: motion modeling based on a linear time series model; establish a p-order autoregressive model (AR) including; a p-order autoregressive model (AR) AR (p) has the following parameters: coefficient matrix Ak∈Rm×m, the parameter v introduced to ensure that the mean value of the dynamic process is non-zero, and the covariance matrix Q of Gaussian white noise. Let Ak be a diagonal matrix, then z(t) Each component of is independent; Given two autoregressive models (AR) A=[v, A1, A2,..., Ap] and A'=[v', A1', A2',..., Ap ’], then the distance metric D(A, A’)=‖A-A′∥F, where ‖·‖F represents the F-norm of the matrix.