CMU's Openpose for ROS

I re-implemented cmu's openpose in tensorflow with some modifications.

Especially, using Mobilenet's 'Depthwise Separable Convolution', I improved it to run in realtime even in an low-computation embedded deivce or only-cpu environment.

See : https://github.com/ildoonet/tf-openpose

• Broadcasting Ros Message
• Humans Pose Estimation
• Face Landmark
• Hand Pose Estimation

See Openpose.

After building openpose, set the environment variable OPENPOSE_HOME.

$ export OPENPOSE_HOME='{path/to/your/openpose/home}'

$ sudo apt-get install ros-indigo-image-common ros-indigo-vision-opencv ros-indigo-video-stream-opencv ros-indigo-image-view

$ cd src
$ git clone https://github.com/ildoonet/ros-openpose
$ cd ..
$ catkin_make

• resolution(string, default: 640x480) : input image resolution
• net_resolution(string, default: 640x480) : network resolution
• num_gpu(int, default: -1) : number of gpus to use. -1 indiciates 'all'.
• num_gpu_start(int, default: 0) : first index of gpu to use.
• model_pose(string, default: COCO) : pose estimation model name provided by openpose.
• no_display(bool, default: false) : if true, it will launch opencv image view to show realtime estimation.
• camera(string, default: /camera/image)
• result_image_topic(string, default: '') : if provided, it will broadcast image with estimated pose.

• /openpose/pose : Msg
  • Persons
    • BodyPartDetection

BodyParts are stored as indexed Nose = 0 Neck = 1 RShoulder = 2 RElbow = 3 RWrist = 4 LShoulder = 5 LElbow = 6 LWrist = 7 RHip = 8 RKnee = 9 RAnkle = 10 LHip = 11 LKnee = 12 LAnkle = 13 REye = 14 LEye = 15 REar = 16 LEar = 17

<node name="openpose_ros_node" pkg="openpose_ros_node" type="openpose_ros_node" output="screen" required="true">
        <param name="camera" value="/videofile/image_raw" />
        <param name="result_image_topic" value="/openpose/image_raw" />
        <param name="resolution" value="480x320" />
        <param name="face" value="false" />
    </node>

$ roslaunch openpose_ros_node videostream.launch video:=0 video_visualize:=true
$ roslaunch openpose_ros_node videostream.launch video:=${filepath} video_visualize:=true

See /launch/videostream.