作者：隋曉金

收到英特爾的開(kāi)發(fā)板-小挪吒，正好手中也有oak相機(jī)，反正都是 OpenVINO一套玩意，進(jìn)行評(píng)測(cè)一下，竟然默認(rèn)是個(gè)Windows系統(tǒng)，刷機(jī)成Linux系統(tǒng)比較方便。

我們先刷個(gè)刷成Linux系統(tǒng)，測(cè)試比較方便，雖然Windows+Python代碼也可以開(kāi)發(fā)，搞點(diǎn)難度的Ubuntu+&++推理，同時(shí)還為了測(cè)試灰仔的ncnn，勉為其難，把系統(tǒng)刷掉，系統(tǒng)我們選擇英特爾適配的22.04即可，確保和CPU的型號(hào)相同即可：

使用motrix的下載，速度較快。然后使用rufus進(jìn)行刻錄優(yōu)盤進(jìn)行sd卡刻入，系統(tǒng)變成linux，就可以遠(yuǎn)程設(shè)置一ssh；系統(tǒng)界面如上。

系統(tǒng)需要安裝官方的OpenVINO組件，使用英特爾端進(jìn)行OpenVINO模型推理，當(dāng)然也可使用ncnn/mnn/onnx，但原聲組件更友好一些。

先配置oak的環(huán)境，適配深度相機(jī)推理和測(cè)距，然后在開(kāi)發(fā)板上推理關(guān)鍵點(diǎn)檢測(cè)推理，演繹一下測(cè)試開(kāi)發(fā)版性能，正好相機(jī)端的芯片也是英特爾使用OpenVINO框架，下面操作是開(kāi)發(fā)板上配置一下相機(jī)使用的庫(kù)環(huán)境：

ubuntu@ubuntu:~$ wget https://gitee.com/oakchina/depthai-core/releases/download/v2.28.0/depthai_2.28.0_amd64.deb
ubuntu@ubuntu:~$ sudo apt install -f
ubuntu@ubuntu:~$ sudo dpkg -i depthai_2.28.0_amd64.deb
(Reading database ... 164136 files and directories currently installed.)
Preparing to unpack depthai_2.28.0_amd64.deb ...
Unpacking depthai (2.28.0) over (2.28.0) ...
Setting up depthai (2.28.0) ...;

配置一下OpenVINO ，參考手冊(cè)。這個(gè)主要后面寫代碼和轉(zhuǎn)模型用。但是我用C++寫代碼，搞點(diǎn)有難度的事情。

https://www.intel.com/content/www/us/en/developer/tools/openvino-toolkit/download.html?PACKAGE=OPENVINO_BASE&VERSION=v_2022_3_2&ENVIRONMENT=DEV_TOOLS&OP_SYSTEM=LINUX&DISTRIBUTION=PIP;

鏈接，下面操作仍然在開(kāi)發(fā)板上執(zhí)行：

pip install openvino-dev==2022.3.2
storage.openvinotoolkit.org


ubuntu@ubuntu:~$ wget https://storage.openvinotoolkit.org/repositories/openvino/packages/2023.3/linux/l_openvino_toolkit_ubuntu22_2023.3.0.13775.ceeafaf64f3_x86_64.tgz
ubuntu@ubuntu:~$ sudo tar xf l_openvino_toolkit_ubuntu22_2023.0.0.10926.b4452d56304_x86_64.tgz.tgz.sha256.tgz -C /opt/intel/
ubuntu@ubuntu:~$ tar -zxvf l_openvino_toolkit_ubuntu22_2023.3.0.13775.ceeafaf64f3_x86_64.tgz
ubuntu@ubuntu:~$ mv l_openvino_toolkit_ubuntu22_2023.3.0.13775.ceeafaf64f3_x86_64 openvino_2023
ubuntu@ubuntu:~$ mv openvino_2023/ /opt/intel/
ubuntu@ubuntu:~$ cd /opt/intel/
ubuntu@ubuntu:~$ cd openvino_2023/
ubuntu@ubuntu:/opt/intel/openvino_2023$ vim ~/.bashrc
source /opt/intel/openvino_2023/setupvars.sh
ubuntu@ubuntu:~$ cd /opt/intel/openvino_2023/install_dependencies/
ubuntu@ubuntu:/opt/intel/openvino_2023/install_dependencies$ sudo -E ./install_openvino_dependencies.sh

下面操作在自己的宿主機(jī)器上執(zhí)行，主要發(fā)現(xiàn)在開(kāi)發(fā)板上的OpenVINO無(wú)法轉(zhuǎn)相機(jī)的blob模型，但是這個(gè)低版本的OpenVINO庫(kù)又無(wú)法開(kāi)發(fā)板，應(yīng)為2021.4支持系統(tǒng)ubuntu20.04版本和一下，開(kāi)發(fā)板的版本是22.04系統(tǒng)版本過(guò)高。

先搞一下yolov5lite，這個(gè)官方給了方法和例子，簡(jiǎn)要敘述和附上，這我是在自己的宿主主機(jī)上做的ubuntu20.04 因?yàn)楝F(xiàn)在開(kāi)發(fā)板版本過(guò)高，擔(dān)心它的OpenVINO環(huán)境轉(zhuǎn)的blob不一定能在oak相機(jī)上運(yùn)行。

ubuntu@ubuntu:~/Downloads$ axel -n 100 https://registrationcenter-download.intel.com/akdlm/IRC_NAS/18096/l_openvino_toolkit_p_2021.4.689.tgz
ubuntu@ubuntu:~$ tar -zxvf l_openvino_toolkit_p_2021.4.689.tgz 
ubuntu@ubuntu:~/Downloads$ cd l_openvino_toolkit_p_2021.4.689/
ubuntu@ubuntu:~/Downloads/l_openvino_toolkit_p_2021.4.689$ sudo ./install_GUI.sh 
ubuntu@ubuntu:~$ cd /opt/intel/openvino_2021/install_dependencies/
ubuntu@ubuntu:/opt/intel/openvino_2021/install_dependencies$ sudo -E ./install_openvino_dependencies.sh 
ubuntu@ubuntu:/opt/intel/openvino_2021/bin$ sudo vim ~/.bashrc 

在末尾添加：

source /opt/intel/openvino_2021/bin/setupvars.sh
ubuntu@ubuntu:/opt/intel/openvino_2021/bin$ source ~/.bashrc 
[setupvars.sh] OpenVINO environment initialized
ubuntu@ubuntu:/opt/intel/openvino_2021/bin$ cd /opt//intel/openvino_2021/deployment_tools/model_optimizer//install_prerequisites/
ubuntu@ubuntu:/opt/intel/openvino_2021/deployment_tools/model_optimizer/install_prerequisites$ sudo ./install_prerequisites.sh

下載模型，進(jìn)行轉(zhuǎn)模型：

ubuntu@ubuntu:~$ git clone https://github.com/ppogg/YOLOv5-Lite

模型代碼，參考o(jì)ak官方代碼：

轉(zhuǎn)onnx模型和轉(zhuǎn)OpenVINO模型 export_onnx.py見(jiàn)官方參考：

ubuntu@ubuntu:~/YOLOv5-Lite$ pip3 install -r requirements.txt
ubuntu@ubuntu:~/YOLOv5-Lite$ python3 export_onnx.py -w v5lite-e.pt -imgsz 640
Namespace(blob=False, convert_tool='blobconverter', img_size=[640, 640], 
input_model=PosixPath('/home/ubuntu/YOLOv5-Lite/v5lite-e.pt'), name='v5lite-e', 
opset=12, output_dir=PosixPath('/home/ubuntu/YOLOv5-Lite'), shaves=6, 
spatial_detection=False)
[18:12:38] INFO   YOLOv5  v1.5-16-g9d649a6 torch 2.4.1+cu121 CPU      
                                        
Fusing layers... 
[18:12:41] INFO   Model Summary: 167 layers, 781205 parameters, 0 gradients, 
          2.9 GFLOPS                         
 
      INFO   Starting ONNX export with onnx 1.16.1...          
      INFO   Starting to simplify ONNX...                
      INFO   ONNX export success, saved as:               
              /home/ubuntu/YOLOv5-Lite/v5lite-e.onnx       
      INFO   anchors:                          
              [10.0, 13.0, 16.0, 30.0, 33.0, 23.0, 30.0, 61.0,  
          62.0, 45.0, 59.0, 119.0, 116.0, 90.0, 156.0, 198.0, 373.0, 
          326.0]                           
      INFO   anchor_masks:                        
              {'side80': [0, 1, 2], 'side40': [3, 4, 5], 'side20':
          [6, 7, 8]}                         
      INFO   Anchors data export success, saved as:           
              /home/ubuntu/YOLOv5-Lite/v5lite-e.json       
      INFO   Export complete (3.61s).                  
ubuntu@ubuntu:~/YOLOv5-Lite$ python3 /opt/intel/openvino_2021/deployment_tools/model_optimizer/mo.py --input_model v5lite-e.onnx --output_dir /home/ubuntu/YOLOv5-Lite/saved/FP16 --input_shape [1,3,640,640] --data_type FP16 --scale_values [255.0,255.0,255.0] --mean_values [0,0,0]
Model Optimizer arguments:
Common parameters:
  - Path to the Input Model:   /home/ubuntu/YOLOv5-Lite/v5lite-e.onnx
  - Path for generated IR:   /home/ubuntu/YOLOv5-Lite/saved/FP16
  - IR output name:   v5lite-e
  - Log level:   ERROR
  - Batch:   Not specified, inherited from the model
  - Input layers:   Not specified, inherited from the model
  - Output layers:   Not specified, inherited from the model
  - Input shapes:   [1,3,640,640]
  - Mean values:   [0,0,0]
  - Scale values:   [255.0,255.0,255.0]
  - Scale factor:   Not specified
  - Precision of IR:   FP16
  - Enable fusing:   True
  - Enable grouped convolutions fusing:   True
  - Move mean values to preprocess section:   None
  - Reverse input channels:   False
ONNX specific parameters:
  - Inference Engine found in:   /opt/intel/openvino_2021/python/python3.8/openvino
Inference Engine version:   2021.4.1-3926-14e67d86634-releases/2021/4
Model Optimizer version:   2021.4.1-3926-14e67d86634-releases/2021/4
[ WARNING ] 
Detected not satisfied dependencies:
  networkx: installed: 3.1, required: ~= 2.5
  numpy: installed: 1.23.5, required: < 1.20
 
Please install required versions of components or use install_prerequisites script
/opt/intel/openvino_2021.4.689/deployment_tools/model_optimizer/install_prerequisites/install_prerequisites_onnx.sh
Note that install_prerequisites scripts may install additional components.
/opt/intel/openvino_2021/deployment_tools/model_optimizer/extensions/front/onnx/parameter_ext.py:20: DeprecationWarning: `mapping.TENSOR_TYPE_TO_NP_TYPE` is now deprecated and will be removed in a future release.To silence this warning, please use `helper.tensor_dtype_to_np_dtype` instead.
 ?'data_type': TENSOR_TYPE_TO_NP_TYPE[t_type.elem_type]
/opt/intel/openvino_2021/deployment_tools/model_optimizer/extensions/analysis/boolean_input.py:13: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
 ?nodes = graph.get_op_nodes(op='Parameter', data_type=np.bool)
/opt/intel/openvino_2021/deployment_tools/model_optimizer/mo/front/common/partial_infer/concat.py:36: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
 ?mask = np.zeros_like(shape, dtype=np.bool)
[ WARNING ] ?Const node '/model.8/Resize/Add_input_port_1/value338417277' returns shape values of 'float64' type but it must be integer or float32. During Elementwise type inference will attempt to cast to float32
[ WARNING ] ?Const node '/model.12/Resize/Add_input_port_1/value341817280' returns shape values of 'float64' type but it must be integer or float32. During Elementwise type inference will attempt to cast to float32
[ WARNING ] ?Changing Const node '/model.8/Resize/Add_input_port_1/value338418006' data type from float16 to  for Elementwise operation
[ WARNING ] Changing Const node '/model.12/Resize/Add_input_port_1/value341817580' data type from float16 to  for Elementwise operation
[ SUCCESS ] Generated IR version 10 model.
[ SUCCESS ] XML file: /home/ubuntu/YOLOv5-Lite/saved/FP16/v5lite-e.xml
[ SUCCESS ] BIN file: /home/ubuntu/YOLOv5-Lite/saved/FP16/v5lite-e.bin
[ SUCCESS ] Total execution time: 10.69 seconds. 
[ SUCCESS ] Memory consumed: 104 MB. 
It's been a while, check for a new version of Intel(R) Distribution of OpenVINO(TM) toolkit here https://software.intel.com/content/www/us/en/develop/tools/openvino-toolkit/download.html?cid=other&source=prod&campid=ww_2021_bu_IOTG_OpenVINO-2021-4-LTS&content=upg_all&medium=organic or on the GitHub*
ubuntu@ubuntu:~/YOLOv5-Lite$ 

轉(zhuǎn)換模型

ubuntu@ubuntu:~$ find . -name "mo_onnx.py"
./.local/lib/python3.10/site-packages/openvino/tools/mo/mo_onnx.py
ubuntu@ubuntu:~$ python3 ./.local/lib/python3.10/site-packages/openvino/tools/mo/mo_onnx.py --input_model v5lite-e.onnx --output_dir /home/ubuntu/YOLOv5-Lite/saved/FP16 --input_shape [1,3,640,640] --data_type FP16 --scale_values [255.0,255.0,255.0] --mean_values [0,0,0]
[ WARNING ] Use of deprecated cli option --data_type detected. Option use in the following releases will be fatal.
Check for a new version of Intel(R) Distribution of OpenVINO(TM) toolkit here https://software.intel.com/content/www/us/en/develop/tools/openvino-toolkit/download.html?cid=other&source=prod&campid=ww_2023_bu_IOTG_OpenVINO-2022-3&content=upg_all&medium=organic or on https://github.com/openvinotoolkit/openvino
[ INFO ] The model was converted to IR v11, the latest model format that corresponds to the source DL framework input/output format. While IR v11 is backwards compatible with OpenVINO Inference Engine API v1.0, please use API v2.0 (as of 2022.1) to take advantage of the latest improvements in IR v11.
Find more information about API v2.0 and IR v11 at https://docs.openvino.ai/latest/openvino_2_0_transition_guide.html
[ SUCCESS ] Generated IR version 11 model.
[ SUCCESS ] XML file: /home/ubuntu/YOLOv5-Lite/saved/FP16/v5lite-e.xml
[ SUCCESS ] BIN file: /home/ubuntu/YOLOv5-Lite/saved/FP16/v5lite-e.bin
ubuntu@ubuntu:~$ pip3 install blobconverter
然后站blob


[setupvars.sh] OpenVINO environment initialized
ubuntu@ubuntu:~/YOLOv5-Lite$ cd /opt/intel/openvino_2021/deployment_tools/tools
ubuntu@ubuntu:/opt/intel/openvino_2021/deployment_tools/tools$ sudo chmod 777 compile_tool/
[sudo] password for ubuntu: 
ubuntu@ubuntu:/opt/intel/openvino_2021/deployment_tools/tools$ cd compile_tool/
ubuntu@ubuntu:/opt/intel/openvino_2021/deployment_tools/tools/compile_tool$ ./compile_tool -m /home/ubuntu/YOLOv5-Lite/saved/FP16/v5lite-e.xml -ip U8 -d MYRIAD -VPU_NUMBER_OF_SHAVES 4 -VPU_NUMBER_OF_CMX_SLICES 4
Inference Engine: 
  IE version ......... 2021.4.1
  Build ........... 2021.4.1-3926-14e67d86634-releases/2021/4
 
Network inputs:
  images : U8 / NCHW
Network outputs:
  output1_yolov5 : FP16 / NCHW
  output2_yolov5 : FP16 / NCHW
  output3_yolov5 : FP16 / NCHW
[Warning][VPU][Config] Deprecated option was used : VPU_MYRIAD_PLATFORM
Done. LoadNetwork time elapsed: 6529 ms
ubuntu@ubuntu:/opt/intel/openvino_2021/deployment_tools/tools/compile_tool$ ls
compile_tool README.md v5lite-e.blob

導(dǎo)出模型，先在oak相機(jī)上試試，這個(gè)整個(gè)模型都是在oak相機(jī)端進(jìn)行推理和測(cè)距，只能說(shuō)這個(gè)開(kāi)發(fā)板是支持oak這種深度相機(jī)使用的。

接著，來(lái)修改我們的代碼，將模型放在開(kāi)發(fā)板上使用OpenVINO推理，將測(cè)距功能仍然保持相機(jī)端推理，下面是使用clion遠(yuǎn)程調(diào)用開(kāi)發(fā)板進(jìn)行編譯的代碼，將深度相機(jī)OAK插在哪吒開(kāi)發(fā)板的usb接口，將英特爾開(kāi)發(fā)板插上顯示器，然后進(jìn)行相機(jī)調(diào)用，后續(xù)上傳GitHub。

cmakelists.txt

cmake_minimum_required(VERSION 3.16)
project(demo)
set(CMAKE_CXX_STANDARD 11)
find_package(OpenCV REQUIRED)
#message(STATUS ${OpenCV_INCLUDE_DIRS})
#添加頭文件
include_directories(${OpenCV_INCLUDE_DIRS})
include_directories(${CMAKE_SOURCE_DIR}/include)
include_directories(${CMAKE_SOURCE_DIR}/include/utility)
#鏈接Opencv庫(kù)
find_package(depthai CONFIG REQUIRED)
add_executable(demo main.cpp include/utility/utility.cpp)
target_link_libraries(demo ${OpenCV_LIBS} depthai::opencv )
 

main.cpp

#include 
// Includes common necessary includes for development using depthai library
#include "depthai/depthai.hpp"
 
/*
The code is the same as for Tiny-yolo-V3, the only difference is the blob file.
The blob was compiled following this tutorial: https://github.com/TNTWEN/OpenVINO-YOLOV4
*/
 
 
static const std::vector labelMap = {
        "person",        "bicycle",      "car",           "motorbike",     "aeroplane",   "bus",         "train",       "truck",        "boat",
        "traffic light", "fire hydrant", "stop sign",     "parking meter", "bench",       "bird",        "cat",         "dog",          "horse",
        "sheep",         "cow",          "elephant",      "bear",          "zebra",       "giraffe",     "backpack",    "umbrella",     "handbag",
        "tie",           "suitcase",     "frisbee",       "skis",          "snowboard",   "sports ball", "kite",        "baseball bat", "baseball glove",
        "skateboard",    "surfboard",    "tennis racket", "bottle",        "wine glass",  "cup",         "fork",        "knife",        "spoon",
        "bowl",          "banana",       "apple",         "sandwich",      "orange",      "broccoli",    "carrot",      "hot dog",      "pizza",
        "donut",         "cake",         "chair",         "sofa",          "pottedplant", "bed",         "diningtable", "toilet",       "tvmonitor",
        "laptop",        "mouse",        "remote",        "keyboard",      "cell phone",  "microwave",   "oven",        "toaster",      "sink",
        "refrigerator",  "book",         "clock",         "vase",          "scissors",    "teddy bear",  "hair drier",  "toothbrush"};
 
static std::atomic syncNN{true};
 
 
int main() {
    // Create pipeline
    dai::Pipeline pipeline;
 
    // Define sources
    auto camRgb = pipeline.create();
    auto monoLeft = pipeline.create();
    auto monoRight = pipeline.create();
    auto stereo = pipeline.create();
    auto spatialDataCalculator = pipeline.create();
 
 
    // Properties
    camRgb->setPreviewSize(640, 640);
    camRgb->setBoardSocket(dai::RGB);
    camRgb->setResolution(dai::THE_1080_P);
    camRgb->setInterleaved(false);
    camRgb->setColorOrder(dai::RGB);
    camRgb->setPreviewKeepAspectRatio(false); //將調(diào)整視頻大小以適應(yīng)預(yù)覽大小，對(duì)齊
 
    monoLeft->setBoardSocket(dai::LEFT);
    monoLeft->setResolution(dai::THE_720_P);
    monoRight->setBoardSocket(dai::RIGHT);
    monoRight->setResolution(dai::THE_720_P);
 
 
    stereo->setDefaultProfilePreset(dai::HIGH_ACCURACY);
    stereo->setLeftRightCheck(true);
    stereo->setDepthAlign(dai::RGB);
    stereo->setExtendedDisparity(true);
 
    dai::Point2f topLeft(0.4f, 0.4f);
    dai::Point2f bottomRight(0.6f, 0.6f);
 
    dai::SpatialLocationCalculatorConfigData config;
    config.depthThresholds.lowerThreshold = 100;
    config.depthThresholds.upperThreshold = 10000;
    config.roi = dai::Rect(topLeft, bottomRight);
 
    spatialDataCalculator->initialConfig.addROI(config);
    spatialDataCalculator->inputConfig.setWaitForMessage(false);
 
 
    // Network specific settings
    auto detectionNetwork = pipeline.create();
    detectionNetwork->setBlob("../v5lite-e.blob");
    detectionNetwork->setConfidenceThreshold(0.5);
    //Yolo specific parameters
    detectionNetwork->setNumClasses(80);
    detectionNetwork->setCoordinateSize(4);
    detectionNetwork->setAnchors({10,13,16,30,33,23,30,61,62,45,59,119,116,90,156,198,373,326});
    detectionNetwork->setAnchorMasks({{{"side80",{0, 1, 2}},{"side40",{3, 4, 5}},{"side20",{6, 7, 8}}}});
    detectionNetwork->setIouThreshold(0.5);
 
    // rgb輸出
    auto xoutRgb = pipeline.create();
    xoutRgb->setStreamName("rgb");
 
    // depth輸出
    auto xoutDepth = pipeline.create();
    xoutDepth->setStreamName("depth");
 
    // 測(cè)距模塊數(shù)據(jù)輸出
    auto xoutSpatialData = pipeline.create();
    xoutSpatialData->setStreamName("spatialData");
 
    // 測(cè)距模塊配置輸入
    auto xinSpatialCalcConfig = pipeline.create();
    xinSpatialCalcConfig->setStreamName("spatialCalcConfig");
 
 
    // Linking  preview 畫布 video 實(shí)時(shí)分辨率
    camRgb->video.link(xoutRgb->input); //顯示用video
    camRgb->preview.link(detectionNetwork->input); //推理用preview
    monoLeft->out.link(stereo->left);
    monoRight->out.link(stereo->right);
 
    spatialDataCalculator->passthroughDepth.link(xoutDepth->input);
    stereo->depth.link(spatialDataCalculator->inputDepth);
 
    spatialDataCalculator->out.link(xoutSpatialData->input);
    xinSpatialCalcConfig->out.link(spatialDataCalculator->inputConfig);
 
 
    // output
    auto xlinkParseOut = pipeline.create();
    xlinkParseOut->setStreamName("parseOut");
 
    auto xlinkoutOut = pipeline.create();
    xlinkoutOut->setStreamName("out");
 
    auto xlinkPassthroughOut = pipeline.create();
    xlinkPassthroughOut->setStreamName("passthrough");
 
 
    detectionNetwork->out.link(xlinkParseOut->input);
    detectionNetwork->passthrough.link(xlinkPassthroughOut->input);
 
 
    // Connect to device and start pipeline
    dai::Device device;
 
    device.setIrLaserDotProjectorBrightness(1000);
    device.setIrFloodLightBrightness(0);
    device.startPipeline(pipeline);
 
    // Output queues will be used to get the rgb frames and nn data from the outputs defined above
    auto detectQueue = device.getOutputQueue("parseOut",8,false);
    auto passthQueue = device.getOutputQueue("passthrough", 8, false);
    auto depthQueue = device.getOutputQueue("depth", 8, false);
    auto spatialCalcQueue = device.getOutputQueue("spatialData", 8, false);
    auto spatialCalcConfigInQueue = device.getInputQueue("spatialCalcConfig", 8, false);
    auto rgbQueue = device.getOutputQueue("rgb", 8, false);
 
    bool printOutputLayersOnce = true;
    auto color = cv::Scalar(0,255,0);
 
 
    std::vector detections;
    auto startTime = std::now();
    int counter = 0;
    float fps = 0;
    auto color2 = cv::Scalar(255, 255, 255);
    cv::Scalar color1 = cv::Scalar(0, 0, 255);
 
    while (true) {
        counter++;
        auto currentTime = std::now();
        auto elapsed = std::duration_cast>(currentTime - startTime);
        if(elapsed > std::seconds(1)) {
            fps = counter / elapsed.count();
            counter = 0;
            startTime = currentTime;
        }
 
        std::shared_ptr inRgb = rgbQueue->get();
        std::shared_ptr inDepth = depthQueue->get();
        std::shared_ptr inDet = detectQueue->get();
        std::shared_ptr ImgFrame = passthQueue->get();
 
        cv::Mat frame = inRgb->getCvFrame();
        cv::Mat src = ImgFrame->getCvFrame();
 
        cv::Mat depthFrameColor;
        cv::Mat depthFrame = inDepth->getFrame();
        cv::normalize(depthFrame, depthFrameColor, 255, 0, cv::NORM_INF, CV_8UC1);
        cv::equalizeHist(depthFrameColor, depthFrameColor);
        cv::applyColorMap(depthFrameColor, depthFrameColor, cv::COLORMAP_HOT);
 
        inDet = detectQueue->get();
        if(inDet) {
            detections = inDet->detections;
            for(auto& detection : detections) {
                int x1 = detection.xmin * src.cols;
                int y1 = detection.ymin * src.rows;
                int x2 = detection.xmax * src.cols;
                int y2 = detection.ymax * src.rows;
 
                uint32_t labelIndex = detection.label;
                std::string labelStr = std::to_string(labelIndex);
                if(labelIndex < labelMap.size()) {
                    labelStr = labelMap[labelIndex];
                }
                cv::putText(src, labelStr, cv::Point(x1 + 10, y1 + 20), cv::FONT_HERSHEY_TRIPLEX, 0.5, 255);
                std::stringstream confStr;
                confStr << std::fixed << std::setprecision(2) << detection.confidence * 100;
                cv::putText(src, confStr.str(), cv::Point(x1 + 10, y1 + 40), cv::FONT_HERSHEY_TRIPLEX, 0.5, 255);
                cv::rectangle(src, cv::Point(x1, y1), cv::Point(x2, y2)), color, cv::FONT_HERSHEY_SIMPLEX);
 
                // 1920*1080
                //cv::rectangle(depthFrameColor, cv::Point(x1, y1), cv::Point(x2, y2)), color, cv::FONT_HERSHEY_SIMPLEX);
                int top_left_x = detection.xmin * frame.cols;
                int top_left_y = detection.ymin * frame.rows;
                int bottom_right_x = detection.xmax * frame.cols;
                int bottom_right_y = detection.ymax * frame.rows;
 
                // 最值限定
                top_left_x = top_left_x < 0 ? 0 : top_left_x;
                bottom_right_x = bottom_right_x > frame.cols - 1 ? frame.cols - 1 : bottom_right_x;
                top_left_y = top_left_y < 0 ? 0 : top_left_y;
                bottom_right_y = bottom_right_y > frame.rows - 1 ? frame.rows - 1 : bottom_right_y;
 
                topLeft.x = top_left_x;
                topLeft.y = top_left_y;
                bottomRight.x = bottom_right_x;
                bottomRight.y = bottom_right_y;
 
                // 測(cè)距模塊推送實(shí)際像素大小的ROI
                config.roi = dai::Rect(topLeft, bottomRight);
                dai::SpatialLocationCalculatorConfig cfg;
                cfg.addROI(config);
                spatialCalcConfigInQueue->send(cfg);
                std::vector spatialData = spatialCalcQueue->get()->getSpatialLocations();
 
                for (auto &depthData : spatialData) {
                    auto roi = depthData.config.roi;
                    roi = roi.denormalize(depthFrameColor.cols, depthFrameColor.rows);
                    auto xmin = (int) roi.topLeft().x;
                    auto ymin = (int) roi.topLeft().y;
                    auto xmax = (int) roi.bottomRight().x;
                    auto ymax = (int) roi.bottomRight().y;
 
                    // 最值限定
//                    xmin = xmin < 0 ? 0 : xmin;
//                    xmax = xmax > frame.cols - 1 ? frame.cols - 1 : xmax;
//                    ymin = ymin < 0 ? 0 : ymin;
//                    ymax = ymax > frame.rows - 1 ? frame.rows - 1 : ymax;
 
                    auto coords = depthData.spatialCoordinates;
                    auto distance = std::sqrt(coords.x * coords.x + coords.y * coords.y + coords.z * coords.z);
                    auto fontType = cv::FONT_HERSHEY_TRIPLEX;
 
                    std::stringstream rgb_depthX, depthX, rgb_depthX_;
                    rgb_depthX << "X: " << (int) coords.x << " mm";
                    rgb_depthX_.precision(2);
                    rgb_depthX_ << "dis: " << std::fixed << static_cast(distance) << " mm";
 
                    cv::rectangle(frame,
                                  cv::Point(xmin, ymin), cv::Point(xmax, ymax)),
                                  color,
                                  fontType);
 
                    cv::putText(frame, rgb_depthX_.str(), cv::Point(xmin + 10, ymin - 20),
                                fontType,
                                0.5, color1);
 
                    cv::putText(frame, rgb_depthX.str(), cv::Point(xmin + 10, ymin + 20),
                                fontType,
                                0.5, color1);
                    std::stringstream rgb_depthY, depthY;
                    rgb_depthY << "Y: " << (int) coords.y << " mm";
                    cv::putText(frame, rgb_depthY.str(), cv::Point(xmin + 10, ymin + 35),
                                fontType,
                                0.5, color1);
                    std::stringstream rgb_depthZ, depthZ;
                    rgb_depthZ << "Z: " << (int) coords.z << " mm";
                    cv::putText(frame, rgb_depthZ.str(), cv::Point(xmin + 10, ymin + 50),
                                fontType,
                                0.5, color1);
 
 
                    cv::rectangle(depthFrameColor,
                            cv::Point(xmin, ymin), cv::Point(xmax, ymax)),
                            color,
                            fontType);
                    depthX << "X: " << (int) coords.x << " mm";
                    cv::putText(depthFrameColor, depthX.str(), cv::Point(xmin + 10, ymin + 20),
                                fontType, 0.5, color1);
                    depthY << "Y: " << (int) coords.y << " mm";
                    cv::putText(depthFrameColor, depthY.str(), cv::Point(xmin + 10, ymin + 35),
                                fontType, 0.5, color1);
                    depthZ << "Z: " << (int) coords.z << " mm";
                    cv::putText(depthFrameColor, depthZ.str(), cv::Point(xmin + 10, ymin + 50),
                                fontType, 0.5, color1);
                }
            }
 
            std::stringstream fpsStr;
            fpsStr << "NN fps: " << std::fixed << std::setprecision(2) << fps;
//            printf("fps %f
",fps);
            cv::putText(src, fpsStr.str(), cv::Point(4, 22), cv::FONT_HERSHEY_TRIPLEX, 1,
                        cv::Scalar(0, 255, 0));
            cv::putText(frame, fpsStr.str(), cv::Point(4, 22), cv::FONT_HERSHEY_TRIPLEX, 1,
                        cv::Scalar(0, 255, 0));
 
            // Show the frame
//            cv::imshow("src", src);
            cv::imshow("frame", frame);
            cv::imwrite("frame.jpg", frame);
//            cv::imshow("depth", depthFrameColor);
            int key = cv::waitKey(1);
            if(key == 'q' || key == 'Q' || key == 27) {
                return 0;
            }
        }
    }
}

然后將在相機(jī)端的推理代碼踢掉，使用本地開(kāi)發(fā)板哪吒進(jìn)行推理，然后整體替換OpenVINO推理方式：

（1）先改個(gè)用編解碼的方式獲取相機(jī)，測(cè)距，使用CPU進(jìn)行純h264解碼，純CPU解碼30幀左右，看樣子還行，這小板子的CPU軟解看著還湊合。

cmakelists.txt

cmake_minimum_required(VERSION 3.16)
project(demo)
set(CMAKE_CXX_STANDARD 11)
find_package(OpenCV REQUIRED)
#message(STATUS ${OpenCV_INCLUDE_DIRS})
#添加頭文件
include_directories(${OpenCV_INCLUDE_DIRS})
include_directories(${CMAKE_SOURCE_DIR}/include)
include_directories(${CMAKE_SOURCE_DIR}/include/utility)
#鏈接Opencv庫(kù)
find_package(depthai CONFIG REQUIRED)
add_executable(demo main.cpp include/utility/utility.cpp)
target_link_libraries(demo ${OpenCV_LIBS} depthai::opencv -lavformat -lavcodec -lswscale -lavutil -lz)

main.cpp

#include 
#include 
#include 
#include 
#include 
#include 
#include 
extern "C"
{
#include 
#include 
#include 
#include 
}
 
 
#include "utility.hpp"
 
#include "depthai/depthai.hpp"
 
using namespace std::chrono;
 
int main(int argc, char **argv) {
  dai::Pipeline pipeline;
  //定義
  auto cam = pipeline.create();
  cam->setBoardSocket(dai::RGB);
  cam->setResolution(dai::THE_1080_P);
  cam->setVideoSize(1920, 1080);
  cam->setFps(30);
  auto Encoder = pipeline.create();
  Encoder->setDefaultProfilePreset(cam->getVideoSize(), cam->getFps(),
                   dai::H265_MAIN);
 
 
  cam->video.link(Encoder->input);
 
  auto monoLeft = pipeline.create();
  auto monoRight = pipeline.create();
  auto stereo = pipeline.create();
  auto spatialLocationCalculator = pipeline.create();
 
  auto xoutDepth = pipeline.create();
  auto xoutSpatialData = pipeline.create();
  auto xinSpatialCalcConfig = pipeline.create();
  auto xoutRgb = pipeline.create();
  xoutDepth->setStreamName("depth");
  xoutSpatialData->setStreamName("spatialData");
  xinSpatialCalcConfig->setStreamName("spatialCalcConfig");
  xoutRgb->setStreamName("rgb");
 
  monoLeft->setResolution(dai::THE_400_P);
  monoLeft->setBoardSocket(dai::LEFT);
  monoRight->setResolution(dai::THE_400_P);
  monoRight->setBoardSocket(dai::RIGHT);
 
  stereo->setDefaultProfilePreset(dai::HIGH_ACCURACY);
  stereo->setLeftRightCheck(true);
  stereo->setExtendedDisparity(true);
  spatialLocationCalculator->inputConfig.setWaitForMessage(false);
 
 
  dai::SpatialLocationCalculatorConfigData config;
  config.depthThresholds.lowerThreshold = 200;
  config.depthThresholds.upperThreshold = 10000;
  config.roi = dai::Point2f( 0.1, 0.45), dai::Point2f(( 1) * 0.1, 0.55));
  spatialLocationCalculator->initialConfig.addROI(config);
 
  // Linking
  monoLeft->out.link(stereo->left);
  monoRight->out.link(stereo->right);
 
  spatialLocationCalculator->passthroughDepth.link(xoutDepth->input);
  stereo->depth.link(spatialLocationCalculator->inputDepth);
 
  spatialLocationCalculator->out.link(xoutSpatialData->input);
  xinSpatialCalcConfig->out.link(spatialLocationCalculator->inputConfig);
 
 
  //定義輸出
  auto xlinkoutpreviewOut = pipeline.create();
  xlinkoutpreviewOut->setStreamName("out");
 
  Encoder->bitstream.link(xlinkoutpreviewOut->input);
 
 
  //結(jié)構(gòu)推送相機(jī)
  dai::Device device(pipeline);
  device.setIrLaserDotProjectorBrightness(1000);
 
  //取幀顯示
  auto outqueue = device.getOutputQueue("out", cam->getFps(), false);//maxsize 代表緩沖數(shù)據(jù)
  auto depthQueue = device.getOutputQueue("depth", 4, false);
  auto spatialCalcQueue = device.getOutputQueue("spatialData", 4, false);
 
  //auto videoFile = std::ofstream("video.h265", std::binary);
 
 
  int width = 1920;
  int height = 1080;
  AVCodec *pCodec = avcodec_find_decoder(AV_CODEC_ID_H265);
  AVCodecContext *pCodecCtx = avcodec_alloc_context3(pCodec);
  int ret = avcodec_open2(pCodecCtx, pCodec, NULL);
  if (ret < 0) {//打開(kāi)解碼器
 ? ? ? ?printf("Could not open codec.
");
 ? ? ? ?return -1;
 ? ?}
 ? ?AVFrame *picture = av_frame_alloc();
 ? ?picture->width = width;
  picture->height = height;
  picture->format = AV_PIX_FMT_YUV420P;
  ret = av_frame_get_buffer(picture, 1);
  if (ret < 0) {
 ? ? ? ?printf("av_frame_get_buffer error
");
 ? ? ? ?return -1;
 ? ?}
 ? ?AVFrame *pFrame = av_frame_alloc();
 ? ?pFrame->width = width;
  pFrame->height = height;
  pFrame->format = AV_PIX_FMT_YUV420P;
  ret = av_frame_get_buffer(pFrame, 1);
  if (ret < 0) {
 ? ? ? ?printf("av_frame_get_buffer error
");
 ? ? ? ?return -1;
 ? ?}
 ? ?AVFrame *pFrameRGB = av_frame_alloc();
 ? ?pFrameRGB->width = width;
  pFrameRGB->height = height;
  pFrameRGB->format = AV_PIX_FMT_RGB24;
  ret = av_frame_get_buffer(pFrameRGB, 1);
  if (ret < 0) {
 ? ? ? ?printf("av_frame_get_buffer error
");
 ? ? ? ?return -1;
 ? ?}
 
 
 ? ?int picture_size = av_image_get_buffer_size(AV_PIX_FMT_YUV420P, width, height,
 ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?1);//計(jì)算這個(gè)格式的圖片，需要多少字節(jié)來(lái)存儲(chǔ)
 ? ?uint8_t *out_buff = (uint8_t *) av_malloc(picture_size * sizeof(uint8_t));
 ? ?av_image_fill_arrays(picture->data, picture->linesize, out_buff, AV_PIX_FMT_YUV420P, width,
             height, 1);
  //這個(gè)函數(shù) 是緩存轉(zhuǎn)換格式，可以不用 以為上面已經(jīng)設(shè)置了AV_PIX_FMT_YUV420P
  SwsContext *img_convert_ctx = sws_getContext(width, height, AV_PIX_FMT_YUV420P,
                         width, height, AV_PIX_FMT_RGB24, 4,
                         NULL, NULL, NULL);
  AVPacket *packet = av_packet_alloc();
 
  auto startTime = steady_clock::now();
  int counter = 0;
  float fps = 0;
  auto spatialCalcConfigInQueue = device.getInputQueue("spatialCalcConfig");
  while (true) {
    counter++;
    auto currentTime = steady_clock::now();
    auto elapsed = duration_cast>(currentTime - startTime);
    if (elapsed > seconds(1)) {
      fps = counter / elapsed.count();
      counter = 0;
      startTime = currentTime;
    }
 
 
 
 
    auto h265Packet = outqueue->get();
 
 
    //videoFile.write((char *) (h265Packet->getData().data()), h265Packet->getData().size());
 
    packet->data = (uint8_t *) h265Packet->getData().data();  //這里填入一個(gè)指向完整H264數(shù)據(jù)幀的指針
    packet->size = h265Packet->getData().size();    //這個(gè)填入H265 數(shù)據(jù)幀的大小
    packet->stream_index = 0;
    ret = avcodec_send_packet(pCodecCtx, packet);
    if (ret < 0) {
 ? ? ? ? ? ?printf("avcodec_send_packet 
");
 ? ? ? ? ? ?continue;
 ? ? ? ?}
 ? ? ? ?av_packet_unref(packet);
 ? ? ? ?int got_picture = avcodec_receive_frame(pCodecCtx, pFrame);
 ? ? ? ?av_frame_is_writable(pFrame);
 ? ? ? ?if (got_picture < 0) {
 ? ? ? ? ? ?printf("avcodec_receive_frame 
");
 ? ? ? ? ? ?continue;
 ? ? ? ?}
 
 ? ? ? ?sws_scale(img_convert_ctx, pFrame->data, pFrame->linesize, 0,
         height,
         pFrameRGB->data, pFrameRGB->linesize);
 
 
    cv::Mat mRGB(cv::Size(width, height), CV_8UC3);
    mRGB.data = (unsigned char *) pFrameRGB->data[0];
    cv::Mat mBGR;
    cv::cvtColor(mRGB, mBGR, cv::COLOR_RGB2BGR);
    std::stringstream fpsStr;
    fpsStr << "NN fps: " << std::fixed << std::setprecision(2) << fps;
 ? ? ? ?printf("fps %f
",fps);
 ? ? ? ?cv::putText(mBGR, fpsStr.str(), cv::Point(4, 22), cv::FONT_HERSHEY_TRIPLEX, 0.4,
 ? ? ? ? ? ? ? ? ? ?cv::Scalar(0, 255, 0));
 
 
 ? ? ? ?config.roi = dai::Point2f(3 * 0.1, 0.45), dai::Point2f((3 + 1) * 0.1, 0.55));
 ? ? ? ?dai::SpatialLocationCalculatorConfig cfg;
 ? ? ? ?cfg.addROI(config);
 ? ? ? ?spatialCalcConfigInQueue->send(cfg);
 
    // auto inDepth = depthQueue->get();
    //cv::Mat depthFrame = inDepth->getFrame(); // depthFrame values are in millimeters
 
 
    auto spatialData = spatialCalcQueue->get()->getSpatialLocations();
    for(auto depthData : spatialData) {
      auto roi = depthData.config.roi;
      roi = roi.denormalize(mBGR.cols, mBGR.rows);
 
      auto xmin = static_cast(roi.topLeft().x);
      auto ymin = static_cast(roi.topLeft().y);
      auto xmax = static_cast(roi.bottomRight().x);
      auto ymax = static_cast(roi.bottomRight().y);
 
      auto coords = depthData.spatialCoordinates;
      auto distance = std::sqrt(coords.x * coords.x + coords.y * coords.y + coords.z * coords.z);
      auto color = cv::Scalar(0, 200, 40);
      auto fontType = cv::FONT_HERSHEY_TRIPLEX;
      cv::rectangle(mBGR, cv::Point(xmin, ymin), cv::Point(xmax, ymax)), color);
      std::stringstream depthDistance;
      depthDistance.precision(2);
      depthDistance << std::fixed << static_cast(distance / 1000.0f) << "m";
 ? ? ? ? ? ?cv::putText(mBGR, depthDistance.str(), cv::Point(xmin + 10, ymin + 20), fontType, 0.5, color);
 ? ? ? ?}
 
 
 
 ? ? ? ?cv::imshow("demo", mBGR);
 ? ? ? ?cv::imwrite("demo.jpg",mBGR);
 
 ? ? ? ?cv::waitKey(1);
 
 
 ? ?}
 
 
 ? ?return 0;
}

整個(gè)代碼在哪吒開(kāi)發(fā)板上進(jìn)行解碼，幀率達(dá)到30fps左右，還可以，圖片就不上傳了，大家可以自己評(píng)測(cè)，前提安裝ffmpeg這個(gè)庫(kù)。

（2）v8的模型轉(zhuǎn)換和開(kāi)發(fā)板上推理，這個(gè)地方一定要保證opset=11,如果是14是不可以的，模型轉(zhuǎn)換可以在開(kāi)發(fā)板上轉(zhuǎn)換就行。

ubuntu@ubuntu:~$ pip install ultralytics轉(zhuǎn)換代碼

ubuntu@ubuntu:~$ cat convert_yolov8.py
from ultralytics import YOLO
 
# Load a model
model = YOLO("yolov8n.yaml") # build a new model from scratch
model = YOLO("yolov8n.pt") # load a pretrained model (recommended for training)
 
# Use the model
# model.train(data="coco8.yaml", epochs=3) # train the model
# metrics = model.val() # evaluate model performance on the validation set
results = model("https://ultralytics.com/images/bus.jpg") # predict on an image
path = model.export(format="onnx") # export the model to ONNX format
path = model.export(format="openvino",opset=11) # export the model to ONNX format
cmkelists.txt


cmake_minimum_required(VERSION 3.12)
project(yolov8_openvino_example)
 
set(CMAKE_CXX_STANDARD 14)
 
find_package(OpenCV REQUIRED)
 
include_directories(
  ${OpenCV_INCLUDE_DIRS}
  /opt/intel/openvino_2023/runtime/include
)
 
add_executable(detect 
  main.cc
  inference.cc
)
 
target_link_libraries(detect
  ${OpenCV_LIBS}
   /opt/intel/openvino_2023/runtime/lib/intel64/libopenvino.so
)

測(cè)試代碼使用官方的即可 ultralytics/examples/YOLOv8-OpenVINO-CPP-Inference at main · ultralytics/ultralytics · GitHub

（3）增加板子使用OpenVINO推理+板子CPU/ffmpeg解碼+推流；oak相機(jī)測(cè)距代碼就不添加了。

發(fā)現(xiàn)這個(gè)模型還是比較重，添加到推理端有點(diǎn)小卡，先不加了，先用CPU進(jìn)行編解碼推流吧，測(cè)試目錄和GitHub地址如下，效果圖如下：

拉流設(shè)置命令

github：https://github.com/sxj731533730/OAK_Rtserver.git

參考資料：

[1] OAK相機(jī)如何將yoloV5lite模型轉(zhuǎn)換成blob格式？_oak china yolov5模型轉(zhuǎn)換-CSDN博客

https://blog.csdn.net/oakchina/article/details/129403986

[2]https://github.com/openvinotoolkit/openvino_notebooks/tree/latest/notebooks/pose-estimation-webcam