Toward Fine-Grained Sleeping Activity Recognition: 3d Extension and an Estimation Try on Joint Position of SLP Dataset
Hiroki Kato, Yu Enokibori, Naoto Yoshida, Kenji Mase · 2021
Sleeping posture estimation including joint positions is important for identifying pressure ulcer risk. Since there are occlusion and privacy problems with camera images, we have been studying 2D joint position estimation from sleeping posture pressure images. However, the 2D joint position does not reveal 3d relationships among body parts, such as crossing legs. Thus, toward fine-grained sleeping activity recognition, 3D joint position estimation is required. To study it, we extend the 2D joint-position data of SLP dataset for 3D and then tried to estimate them with high accuracy. In this paper, described the details of the 3D extension and an estimation result. With one network and two loss extensions for a 2D to 3D joint position estimation network, we achieved 6.909 ± 0.278 cm accuracy assuming the average skeleton of Japanese, with about 42.5% error reduction by the extensions.