Abstract: ‍ ‍In different time scales in the process of violence recognition， a multilevel-motion feature violence recognition algorithm based on Lagrange field is proposed in this paper. In this algorithm， the Lagrange field describing the nonlinear particle motion is introduced into the process of violence analysis. The opt Lagrange field based on optical flow is constructed to mine the motion characteristics of violence at different time scales， and a multilevel-motion module based on Lagrange field is designed. The module can calculate the multilevel-motion characteristics according to the length of the input optical flow sequence； Then， a dual flow network based on flow gate control mechanism is constructed to fuse multilevel-motion features and RGB image features； Finally， the recognition results are calculated by using LSTM and full connection model. Experiments show that this method has achieved good results in public violence identification data set， especially in RWF-2000 data set of real monitoring scene， the correct recognition rate of violence identification can reach 88.4%， which is better than other algorithms.