A Human Pose Recognition Model Comparison to Label Radar Data for Gesture Recognition in Robotics

Instantaneous radar data labeling during data set acquisition, through camera based computer vision methods, is a challenging task. In context of human pose and gesture recognition, there is a huge amount of computer vision models available and whether they are suitable for instantaneous data labeling needs to be investigated. Therefore, this paper compares 20 different models regarding inference time and human keypoint location deviations. The most suitable model is statistically investigated further using more than 250 thousand images, where person presence detection and gesture detection are evaluated using simple threshold methods. Obtained results reveal 94% of all images are correctly classified.