The invention discloses a multi-scale Gaussian-Markov random field model-based lower limb motion identification method. The method comprises the steps of 1, collecting original data of C motion modesof human lower limbs, performing preprocessing on the original data, and constructing a data feature graph; 2, performing multi-scale decomposition and feature field modeling of a signal; 3, setting an iterative frequency Q, and iteratively updating a first-layer scale image observation field Gaussian model parameter {mu c, sigma c, pi c}; and 4, performing matching on a classification result anda standard set of each motion mode to judge mode attribution, and according to a voting rule, eliminating the problem of boundary fuzziness of a time sequence segmentation region of different motion modes. According to the method, a stable signal local feature can be extracted, and the signal noise influence caused by motion instability can be eliminated, so that the identification accuracy and the prediction reliability of the lower limb motion modes are improved, and the technical support is provided for stable control of an auxiliary motion system.