Multi-Human Pose Detection Based on EELAN-Blazepose Model

摘要

The human pose estimation system is an exciting topic for developing a collaborative robot. The robot can interpret human poses and autonomously perform collaborative action using various sensors such as cameras and lidars as input devices and use the pose data to interact like giving and receiving objects from humans. There are multiple models for detecting human poses such as Openpose and Mediapipe, but only a few models can detect poses in three dimensions for cases where multiple humans are detected. In this study, we aimed to develop a sstacked model for detecting three-dimensional human poses for cases where multiple humans are detected by combining the EELAN and Blazepose models as our first step in enabling human-robot interactions like giving and taking objects. In the static image tests, our model successfully detected multiple humans and estimated their three-dimensional models separately. On the other hand, the real-time test results showed that our model successfully estimated the three-dimensional human pose for this case with a mean processing time of 354.30 milliseconds(ms) to process 20 frames per batch.