<![CDATA[Human pose estimation is a research topic that has goal to estimate every human’s keypoint coordinate that can be connected and make a human skeleton. The development of this topic can be applicated to human activity recognition, human tracking, and motion capture for film and animation. There are several challenges for this topic: diverse human pose, diverse body appearance from clothing and similar parts, and complex environment that may cause foreground occlusion. There are several methods to be used in this research: YOLOv4- Tiny, EfficientNet Simple Baseline, and VideoPose3D. YOLOv4- Tiny will process image input to get bounding box coordinate. This coordinate will be inputted to EfficientNet Simple Baseline modification to get 16 keypoint 2D coordinates. After that, VideoPose3D will processed 2D coordinates into 15 keypoints 3D coordinates. The result from this research is EfficientNet Simple Baseline modification is faster with 4.54ms time compared to its original with time of 5.15ms. Although faster, its modification has its own downside. In term of accuracy, modification still less accurate than its original with highest average Percentage of Correct Keypoints head (PCKh@0.2) 86.89%, and original with PCKh@0.2 89.62%. This affect 3D human pose estimation using VideoPose3D, where using EfficientNet modification resulting Mean Per Joints Position Error (MPJPE) 25.3 mm compared to original Simple Baseline resulting MPJPE 28.1mm.]]>
Copyrights © 2022
