In order to enable rescue robots to accurately detect human joint points in dangerous building environments to replace rescuers into dangerous scenes. In this paper, we propose a machine learning based human joint point detection method. First, We acquired color images and depth information by kinect camera, obtains human joint point detection data by machine learning, and uses normalization method to smooth the joint point data for pre-processing. Then, we add the human joint length constraint and joint rotation angle constraint to the preprocessed data to generate the spatial position of each human joint point. Finally, the Tsai method is used for hand-eye calibration to obtain the distance between the robotic arm and the casualty. The experimental results showed that the method accurately measures the position of the casualty joint point in the rescue environment of a dangerous building, which verifies the effectiveness of the proposed detection method.