Hi,
Currently, I am working on human skeleton extraction form the depth data by using machine learning or neural network approach. I am not using kinect camera in this case. I am using depth and RGB data stream to extract skeleton. It is very helpful, if you recommend any research papers or method on this.
Thank you,
You can throw any amount of machine learning at it, but the information just is not there. It cannot be done using depth data. Thresholding a CT scan would be more appropriate.
Hi Lambert,
In this case, I am using only stereo camera. So, I am getting depth information and RGB data.
The complexity of the problem deppends pretty much on the quality of your data. Ideally you want as the first step to segment the body from the background. You can use all sorts of pattern recognition thresholding for this issue.
The next step is to transform the segmented object into a skeleton. Here there is also a couple of options, in computer vision this problem is called finding the topological skeleton or skeletonization. Also the concepts of skeleton and medial axis are used interchangeably sometimes.
Also you my find useful the literature on human pose estimation and keywords such as visual hull may probably be relevant.
There is a surprisingly amount of literature on those topics, and used techniques go from using simple morphological operations to complex model matching, segment descriptors, active contours, epipolar geometry ... I recomend you to do a comprehensive review of these literature to try to find some starting point.
Hi Alvaro,
Thank you for your response. I have successfully done the segmentation. I am unable to find the literature for skeletonization from the depth data. Also, direct skeleton data from the kinect camera is used for the human pose estimation. Currently, I am trying different approaches but I am unable to get the perfect results. It is very helpful, if you recommend some research paper on this topic.
Thank you.
Check the following papers. Both paper producing real time results. First paper based on RGB image but i am sure that you can extend this for depth images. Second, paper uses depth image based hand pose estimation which can be easily extend for full body pose estimation.
1) Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields
2) 3D Convolutional Neural Networks for Efficient and Robust Hand Pose Estimation from Single Depth Images
- Badges
- Science topic