The invention discloses a human body posture visual identification method of a transfer carrying nursing robot. According to the method, a two-stage network (the first-stage neural network is PAF andthe second-stage neural network is improved ResNet) based on RGBD (color image RGB + depth image Depth) is utilized to realize close-range higher-precision identification; a ResNet network commonly used for classification is improved, and input and output structures of the ResNet network are modified, so that the ResNet network can solve the problem of human joint recognition, serves as a second-level neural network, and obtains higher precision compared with the existing method. According to the nearest neighbor insect following method, under the condition that only two joint coordinates aredepended on, the armpit contour of the human body is automatically tracked, the armpit contour is further repaired through the convex hull algorithm, the armpit center, namely the armpit point, is finally obtained, and by means of the armpit point recognition method, the dependence of the armpit point on the joint position is reduced, and meanwhile accuracy is improved.