Aiming at the problems of multi-target human body posture estimation

such as low recognition rate and poor robustness caused by body shape difference

posture diversity and occlusion fusion among bodies

a real-time multi-target human body pose estimation algorithm based on depth image is proposed. The method removes the ground by random sampling consensus（RANSAC） algorithm and performs background modeling to segment the human body model. The approximate K-nearest neighbor（ANN） method is used to find the estimated positions of K single human joints from the database created in the learning phase to form the estimated position point cloud. The Mean-Shift algorithm is used to calculate the maximum weighted density point of the point cloud as the estimated joint point. Joint points are connected in a certain order to obtain the human skeleton

and the human body posture is estimated. The algorithm experimental platform is built on Microsoft Visual Studio 2017. The results show that the method can effectively solve the joint point measuring problems caused by self-occlusion and other occlusion.Compared with the existing algorithms on the MPII dataset

the proposed method has average accuracy of 85. 02 %for joint point estimation

and has good robustness.