In view of the high complexity of limb recognition in OpenPose, it was proposed to complete the human skeleton extraction based on TfPose, and the neural network integrated learning method was used to perform limb recognition on the robot lifting instructions to complete the intelligent lifting operation. Firstly, the D-H method was used to perform the forward kinematics analysis of the hoisting robot to determine the working space range of the hoisting mechanism, and the inverse kinematics was solved by the conformal geometric algebra method. The hoisting robot was mathematically modeled from the current position to the target position; then it was obtained based on TfPose. The human skeleton vector and RGB skeleton map were based on BP neural network and InceptionV3 network. The neural network integrated learning method was used to determine the optimal weight to complete the hoisting signal identification. Finally, the identified hoisting command limb signal was transmitted to the hoisting robot through UDP communication. The module was controlled to complete the lifting operation. The experimental results showed that the average limb recognition accuracy of the method was 0977, which solved the large cargo lifting occasions such as ports, docks and mines, and greatly improved the hoisting and loading efficiency.