Estimating 3D Human Pose using Point Based Pose Estimation and Single Stage Method
Uday Kiran G, Vellanki Srilakshmi, Nandhini Ch, Gagan Chandra B, Yamini Keerthi K · 2022
Graphically estimating a person's pose is a computer vision task called human pose estimation. Predicting the location of a person's body parts or joints is an everyday use case for this method. The vast number of potential uses for this kind of technology has made it one of the most intriguing areas of study in computer vision. To estimate a human's 3D stance from 2D photos, the location of the important body joints in 3D space shall be determined. Despite extensive research in the field of computer vision, this issue continues to pose significant difficulties. This is because the depth information is lost when obtaining 2D images of 3D objects, leading to ambiguity. While using a 2D Image as input, only a portion of a person's body is exhibited; estimating their 3D position is challenging. In this study, we offer methods for recovering a human's 3D position from a 2D photograph when just a portion of the body is visible. Any situation where some of the body's major joints are outside the image’s bounds is considered a partial body presence. The presence of these joints in the input image is also detected by a CNN network that we propose. We propose integrating the two networks to determine the full human posture in 3 Dimensional Space and the partial 3D human pose calculated by the detection network. Comparing our experimental results to those of the state-of-the-art demonstrates the efficacy of our methods in the presence of a partially-occluded body. Furthermore, the experimental results show that the 3D human posture may be estimated more accurately from 2D pictures using a linear regression rather than an intermediary stage of 2D pose estimation.