light-OpenPose OpenVINO搭建测试

怼烎@ 2023-10-11 10:28 10阅读 0赞

搭建light openpose openVINO 环境，测试CPU下姿态估计效果。主要参考Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose。  
paper link:https://arxiv.org/abs/1811.12004  
github:https://github.com/Daniil-Osokin/lightweight-human-pose-estimation.pytorch

**测试环境**：

*  Ubuntu 18.04 LTS
 *  i7-7700K CPU @ 4.20GHz
 *  OpenVINO version 2019.2.275
 *  Video: h264 (Main) (avc1 / 0x31637661), yuv420p(tv, bt709, progressive), 640x360 \[SAR 1:1 DAR 16:9\]

**1.python代码测试**  
克隆仓库，下载模型，使用python脚本进行测试。详见 https://github.com/Daniil-Osokin/lightweight-human-pose-estimation.pytorch

**2.安装OpenVINO**  
下载地址：https://software.intel.com/en-us/openvino-toolkit/choose-download  
选择相应版本，下载完成后，执行：

cd ~/Downloads/
    tar -xvzf l_openvino_toolkit_p_<version>.tgz
    cd l_openvino_toolkit_p_<version>
    sudo ./install_GUI.sh

进入图形界面安装，Next… 完成安装。默认安装完成后在/opt/inter/openvino/文件夹下。  
这里提供一个openvino 2019.2.275的百度云下载地址 l\_openvino\_toolkit\_p\_2019.2.275.tgz;

安装依赖：

cd /opt/intel/openvino/install_dependencies
    sudo -E ./install_openvino_dependencies.sh

配置环境：

vim <user_directory>/.bashrc
    source /opt/intel/openvino/bin/setupvars.sh

**3.将PyTorch模型转换为ONNX格式**  
使用下载的仓库中的convert\_to\_onnx.py脚本，进行模型格式转换：

python scripts/convert_to_onnx.py --checkpoint-path <CHECKPOINT>. It produces human-pose-estimation.onnx

**4.将ONNX格式模型转换为OpenVINO格式**  
<OpenVINO\_INSTALL\_DIR>/deployment\_tools/model\_optimizer/mo.py --input\_model human-pose-estimation.onnx --input data --mean\_values data\[128.0,128.0,128.0\] --scale\_values data\[256\] --output stage\_1\_output\_0\_pafs,stage\_1\_output\_1\_heatmaps  
完成这一步骤后会得到human-pose-estimation.xml文件，后面会用到。

**5.build sample**  
human-pose-estimation样例在/opt/intel/openvino\_2019.2.275/inference\_engine/demos/下，在该文件下下执行

bash build_demos.sh

如果提示有未安装的依赖ONNX,test-generator,defusedxml,cmake等安装，即可。

pip install onnx==<required version>
    pip install test-generator==<required version>
    pip install defusedxml==<required version>
    sudo apt-get -y install cmake

执行完成：

[ 95%] Building CXX object smart_classroom_demo/CMakeFiles/smart_classroom_demo.dir/src/image_grabber.cpp.o
    [ 96%] Building CXX object gaze_estimation_demo/CMakeFiles/gaze_estimation_demo.dir/src/utils.cpp.o
    [ 97%] Building CXX object smart_classroom_demo/CMakeFiles/smart_classroom_demo.dir/src/logger.cpp.o
    [ 98%] Building CXX object smart_classroom_demo/CMakeFiles/smart_classroom_demo.dir/src/reid_gallery.cpp.o
    [ 99%] Building CXX object smart_classroom_demo/CMakeFiles/smart_classroom_demo.dir/src/tracker.cpp.o
    [ 99%] Linking CXX executable ../intel64/Release/gaze_estimation_demo
    [ 99%] Built target gaze_estimation_demo
    [100%] Linking CXX executable ../intel64/Release/pedestrian_tracker_demo
    [100%] Built target pedestrian_tracker_demo
    [100%] Linking CXX executable ../intel64/Release/smart_classroom_demo
    [100%] Built target smart_classroom_demo
    Scanning dependencies of target ie_samples
    [100%] Built target ie_samples
    
    Build completed, you can find binaries for all demos in the /home/jerry/omz_demos_build/intel64/Release subfolder.

生成的执行文件在`<user_path>/omz_demos_build/intel64/Release` 文件夹下，然后切换到该文件下执行:

./human_pose_estimation_demo -m <pathTo>/human-pose-estimation.xml -i <pathTo>/yourvideo.mp4

**6.结果**  
![light openpose openvino result1][]

![light openpose openvino result1][light openpose openvino result1 1]

[light openpose openvino result1]: https://imgconvert.csdnimg.cn/aHR0cHM6Ly9ub3RlLnlvdWRhby5jb20veXdzL3B1YmxpYy9yZXNvdXJjZS9hMGQyM2ViZjE3ZmRmN2NlZGQ1YjYxNDkxZThiMWYzMC94bWxub3RlL1dFQlJFU09VUkNFZTdmZTI1NTk0MjQzYzhhMzVhMDM3MzNlNmI1N2M1ODUvMjE5MDE?x-oss-process=image/format,png
[light openpose openvino result1 1]: https://imgconvert.csdnimg.cn/aHR0cHM6Ly9ub3RlLnlvdWRhby5jb20veXdzL3B1YmxpYy9yZXNvdXJjZS9hMGQyM2ViZjE3ZmRmN2NlZGQ1YjYxNDkxZThiMWYzMC94bWxub3RlL1dFQlJFU09VUkNFMGYwZmJjNzc5YTFiNDZkMDlmNTU2OGQxYmRmNjFhMjAvMjE5MDM?x-oss-process=image/format,png