The invention discloses an identification method of human body fine-grained motion on the basis of a UWB (Ultra WideBand) radar and an optimal SVM (Support Vector Machine). A training model established by the identification model is closer to the time frequency characteristic value of a practical motion signal on a distance. On the basis of the optimal SVM of priori location information, an optimal model which has a corresponding distance with data to be predicated can be found, and a problem which is always generated by a conventional overall SVM that misjudgment is caused since different motions in different distances always have similar micro-Doppler characteristics can be favorably avoided. Therefore, by use of the method, identification accuracy can be effectively improved.