ROS package for human pose estimation with Microsoft SimpleBaselines algorithm.

Input to the network should be cropped image of a person (detected person).

You need to download weights for COCO or MPII dataset from following link.

You need usb-cam ROS package for testing it on your PC with webcam.

Launch your cam, in this case web-cam:

roslaunch usb_cam usb_cam-test.launch 

After downloading weights to /home/developer/catkin_ws/src/hpe_ros_package/src/lib/models you can run your HPE inference with following:

roslaunch hpe webcam_inference.launch

Starting procedure for HMI is to setup ROS_MASTER_URI and ROS_IP to the values of host PC to enable streaming to AR glasses. After that we can run following script:

roslaunch hpe hmi_integration.launch 

It's not neccessary to setup ROS_MASTER_URI script anymore because currently stream is being sent using RTSP protocol.

Current transport from server to client currently takes:

If we do not compress image.

average rate: 9.519
	min: 0.041s max: 0.244s std dev: 0.02216s window: 2050

While on raspberry we have subscription frequency 2.5 Hz.

After using image compression with image_transport I get same values.

Launch for using compressed is:

 rosrun image_transport republish raw in:=stickman_cont_area out:=stickman_compressed

Currently it's possible to use human pose estimation for generating references for UAV.

Now we have attitude zone control which sends commands to UAV in joy msg type. It's possible to easily implement position control also.

Available neural networks for HPE which are similar to SimpleBaselines are:

• HRNet
• LPN-Simple and lightweight pose estimation

• Check launch files / Add depth and control types arguments
• Add calibration and dynamic control to arguments in launch files
• Finish depth control (pitch/depth control)
• Try HRNet --> TBD (not neccesary for now)
• [

• Try 3d pose estimation
• Add simple object detection (darknet_ros) for better hpe
• Implement correct post processing as stated here