The research of human-robot collaboration for intelligent manufacturing is being paid gradually increasing attention due to high flexibility and high manufacturing efficiency. Comparing with the traditional manufacturing with low flexibility, human-robot collaboration in manufacturing system provides more personalized and flexible way to cover the shortages of traditional manufacturing mode. In human-robot collaboration system, human motion position prediction in the collaborative space is an essential prerequisite for ensuring the safety of workers. In this paper, 3D sensor Kinect is utilized to directly obtain human joint information. A partial circle delimitation method is used to solve the offset phenomenon of human joint obtained by Kinect, so as to achieve accurate estimation of human joint points. On this basis, an algorithm combing multilayer perceptron and long short-term memory network is explored to predict human motion position accurately. It not only helps to avoid complex feature extraction due to its end-to-end characteristic, but also provide natural interaction manner between human and robot without wearable devices or tags that may become a burden for the former. After that, the experimental results demonstrate that the proposed method makes predicting results accurate, and provides the reliable basis for human position prediction in the human-robot collaboration. This research could be applied to the human motion position prediction in human-robot collaboration process.