 |
Estimate Body Pose Using Deep Learning Using MatLab
This study follows an example of MatLabTM (The MathWorks Inc., USA) [5]. Therefore, the pre-train network is based on the Open Neural Network Exchange (ONNX) which is an open ecosystem for interoperable AI models [6] at first. The pattern of human pose is the MS COCO dataset which is a large-scale object detection, segmentation, key-point detection, and captioning dataset [7]. Then, the tested image was predicted the heatmaps and part affinity fields (PAFs) , which are output from the 2-D output convolutional layers. The post-processing part of the algorithm identifies the individual poses of the people in the image using the heatmaps and PAFs returned by the neural network.
MS COCO human skeleton
The MS COCO data set was proposed in 2014 and is the mainstream in the field of human posture recognition. It focuses on solving large datasets such as object detection, key point detection, object segmentation and subtitle generation in natural environments by using computer vision technology. In 2016, we added a new task to the MS COCO dataset using 2D human skeleton key point detection. The data set contains human samples labeled with coordinates of key points on the human body. These samples are labeled with 17 key points on the human skeleton. Below we combine pictures and specific tables to learn more. Figure 1 shows the 17 key points of the MS COCO human skeleton.
|
IEEE/ICACT20240435 Slide.06
[Big Slide]