Human Pose Estimation In Static Images

The method for human pose estimation mainly included the geometrical relationship model, the observation model and the reasoning algorithm.Because the geometrical relationship model didn’t constrain the relationship of symmetric parts,the results had the missing detection problem in symmetric parts.The traditional method used the HOG feature to establish the observation model.Because the HOG feature didn’t fully express the shape feature of the human part,the observation model could not accurately detect the location of the human part.So the work of this paper was to improve the geometry model and the observation model.Aimed at the missing detection problem in symmetric parts of the human body, the method for human pose estimation based on connection and symmetry was proposed. Angle relationship was used to evaluate the symmetrical degree of the human body’s lower limb. Appearance feature was utilized to evaluate the symmetrical degree of the lower and upper limb, and the connection and symmetry relationship were utilized to estimate human body pose. The experimental results were presented to show that the method improved the accuracy of human pose estimation.Convolutional Neural Networks was used to improve the geometry model.Every part of the body corresponded to a CNN,and the last layer of CNN was mapped to a two- dimensional matrix.Every element of the matrix corresponded to a local area of the input image,and the element’s value representing the area of the body part was used to identify the human part.Compared with the similar method,the proposed method had higher accuracy of pose estimation.