ROS Resources: Documentation | Support | Discussion Forum | Service Status | Q&A answers.ros.org

Openpose Implementation for Mobile / Embedded


#1

Implementation of Openpose for realtime : Human Pose Estimation

This will run on your laptop or embedded devices (even without gpu).

It seems to be a good technique to create various application examples.


[Call for join on opensource project]

I recently re-implemented CMU’s openpose which is one of the best model to estimate human pose in realtime with powerful 'GPUs’
With that reimplementation, I have made a couple of changes like tuning network architecture.

Especially, by using mobilenet’s ‘Depthwise Separable Convolution’, the model can be improved to run realtime even in an low-power embedded device or laptop without a GPU.

But I think, there are much room for improvement in speed and accuracy. So I’m finding crews to help this opensource project to research further. Anyone can join via online.


#2

Anyway you might recommend to extract skeleton points as 3D point coordinates?


#3

Sorry for the late reply. Actually, I’m working on it although It will take some time…


#4

I added ROS support on this project as well as other features.

See https://github.com/ildoonet/tf-pose-estimation/blob/master/etcs/ros.md