OPENCV——ROCKX+RV1126视频流检测人脸

Deitymoon2026-07-03 14:03

一、大体流程图

上图是rockx+rv1126的大体流程，首先要初始化模块包括VI模块、VENC模块、并启动VI模块采集视频流、rockx模块的初始化。初始化模块后，就要分两个线程处理了。

主线程是负责rockx对VI视频流的处理，并用OPENCV对人脸进行画框，最后把处理后的VI数据传输到VENC模块里面。

第二个线程rockx_face_detect_venc_thread，从VENC模块获取到H264的编码码流数据，并把VENC码流数据保存。

二．代码

复制代码 
#include <assert.h>
#include <cstddef>
#include <fcntl.h>
#include <getopt.h>
#include <opencv2/imgproc.hpp>
#include <opencv2/imgproc/imgproc_c.h>
#include <pthread.h>
#include <signal.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <unistd.h>

// #include "common/sample_common.h"
#include "rkmedia_api.h"
#include "rknn_rockx_include/rockx_type.h"
#include "rockx.h"
#include <opencv2/core.hpp>
// #include <opencv2/imgoroc.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/opencv.hpp>

#define CAMERA_PATH "rkispp_scale0"
#define CAMERA_ID 0
#define CAMERA_CHN 0
#define VENC_CHN 0
#define WIDTH 1920
#define HEIGHT 1080
using namespace cv;

void * rockx_face_detect_venc_thread(void * args)
{
    pthread_detach(pthread_self());
    FILE * face_detect_h264 = fopen("face_detect_venc.h264", "w+");

    MEDIA_BUFFER mb = NULL;

    while(1)
    {
        mb = RK_MPI_SYS_GetMediaBuffer(RK_ID_VENC, VENC_CHN, -1);
        if(!mb)
        {
            printf("Get Rockx_Venc Data berek...\n");
            break;
        }
        printf("Get Rockx_Venc Data Success...\n");

        fwrite(RK_MPI_MB_GetPtr(mb), RK_MPI_MB_GetSize(mb), 1, face_detect_h264);
        RK_MPI_MB_ReleaseBuffer(mb);
    }
    return NULL;
}

int main(int argc, char **argv)
{

    int ret;
    VI_CHN_ATTR_S vi_chn_attr;
    vi_chn_attr.pcVideoNode = CAMERA_PATH;        // Path
    vi_chn_attr.u32Width = WIDTH;                  // Width
    vi_chn_attr.u32Height = HEIGHT;                 // Height
    vi_chn_attr.enPixFmt = IMAGE_TYPE_NV12;       // ImageType
    vi_chn_attr.enBufType = VI_CHN_BUF_TYPE_MMAP; // BufType
    vi_chn_attr.u32BufCnt = 3;                    // Cnt
    vi_chn_attr.enWorkMode = VI_WORK_MODE_NORMAL; // Mode
    ret = RK_MPI_VI_SetChnAttr(CAMERA_ID, CAMERA_CHN, &vi_chn_attr);
    if (ret)
    {
        printf("Vi Set Attr Failed.....\n");
        return 0;
    }
    else
    {
        printf("Vi Set Attr Success.....\n");
    }

    ret = RK_MPI_VI_EnableChn(CAMERA_ID, CAMERA_CHN);
    if (ret)
    {
        printf("Vi Enable Attr Failed.....\n");
        return 0;
    }
    else
    {
        printf("Vi Enable Attr Success.....\n");
    }

    VENC_CHN_ATTR_S venc_chn_attr;
    memset(&venc_chn_attr, 0, sizeof(VENC_CHN_ATTR_S));
    venc_chn_attr.stVencAttr.u32PicWidth = WIDTH;
    venc_chn_attr.stVencAttr.u32PicHeight = HEIGHT;
    venc_chn_attr.stVencAttr.u32VirWidth = WIDTH;
    venc_chn_attr.stVencAttr.u32VirHeight = HEIGHT;
    venc_chn_attr.stVencAttr.imageType = IMAGE_TYPE_NV12;
    venc_chn_attr.stVencAttr.enType = RK_CODEC_TYPE_H264;
    venc_chn_attr.stVencAttr.u32Profile = 66;
    venc_chn_attr.stRcAttr.enRcMode = VENC_RC_MODE_H264CBR;
    venc_chn_attr.stRcAttr.stH264Cbr.u32Gop = 25;
    venc_chn_attr.stRcAttr.stH264Cbr.u32BitRate = WIDTH * HEIGHT * 3;
    venc_chn_attr.stRcAttr.stH264Cbr.fr32DstFrameRateDen = 1;
    venc_chn_attr.stRcAttr.stH264Cbr.fr32DstFrameRateNum = 25;
    venc_chn_attr.stRcAttr.stH264Cbr.u32SrcFrameRateDen = 1;
    venc_chn_attr.stRcAttr.stH264Cbr.u32SrcFrameRateNum = 25;
    ret = RK_MPI_VENC_CreateChn(VENC_CHN, &venc_chn_attr);
    if (ret)
    {
        printf("ERROR: Create venc failed!\n");
        exit(0);
    }

    ret = RK_MPI_VI_StartStream(CAMERA_ID, CAMERA_CHN);
    if (ret)
    {
        printf("RK_MPI_VI_StartStream Failed.....\n");
        return 0;
    }
    else
    {
        printf("RK_MPI_VI_StartStream Success.....\n");
    }

    pthread_t pid;
    pthread_create(&pid, NULL, rockx_face_detect_venc_thread, NULL);

    rockx_config_t * face_detect_config = rockx_create_config();
    rockx_add_config(face_detect_config, ROCKX_CONFIG_DATA_PATH, "/userdata/rockx_data");

    rockx_handle_t face_detect_handle;
    rockx_ret_t face_detect_ret;

    rockx_module_t face_detect_module = ROCKX_MODULE_FACE_DETECTION_V2;
    face_detect_ret = rockx_create(&face_detect_handle, face_detect_module, face_detect_config, 0);
    if(face_detect_ret != ROCKX_RET_SUCCESS)
    {
        printf("rockx_creat face_detect failed...\n");
        return -1;
    }
    rockx_release_config(face_detect_config);

    rockx_image_t rv1126_rockx_image;
    memset(&rv1126_rockx_image, 0 ,sizeof(rv1126_rockx_image));
    rv1126_rockx_image.width = WIDTH;
    rv1126_rockx_image.height = HEIGHT;
    rv1126_rockx_image.pixel_format = ROCKX_PIXEL_FORMAT_YUV420SP_NV12;

    MEDIA_BUFFER mb;
    while(1)
    {
        mb = RK_MPI_SYS_GetMediaBuffer(RK_ID_VI, CAMERA_CHN, -1);
        if(!mb)
        {
            printf("Get Vi Stream break....\n");
            break;
        }

        rv1126_rockx_image.data = (uint8_t *)RK_MPI_MB_GetPtr(mb);
        rv1126_rockx_image.size = RK_MPI_MB_GetSize(mb);

        rockx_object_array_t face_detect_array;
        face_detect_ret = rockx_face_detect(face_detect_handle, &rv1126_rockx_image, &face_detect_array, NULL);
        if(face_detect_ret != ROCKX_RET_SUCCESS)
        {
            printf("face_detect failed....\n");
        }

        Mat rv1126_rockx_mat = Mat(HEIGHT, WIDTH, CV_8UC1, rv1126_rockx_image.data);

        for(int i = 0; i < face_detect_array.count; i++)
        {
            int left = face_detect_array.object[i].box.left;
            int top = face_detect_array.object[i].box.top;
            int w = face_detect_array.object[i].box.right - face_detect_array.object[i].box.left;
            int h = face_detect_array.object[i].box.bottom - face_detect_array.object[i].box.top;
            Rect boundingRect(left, top, w, h);
            rectangle(rv1126_rockx_mat, boundingRect, Scalar(255,255,0));
       }
       RK_MPI_SYS_SendMediaBuffer(RK_ID_VENC, VENC_CHN, mb);
       RK_MPI_MB_ReleaseBuffer(mb);
    }

    RK_MPI_VENC_DestroyChn(VENC_CHN);
    RK_MPI_VI_DisableChn(CAMERA_ID, CAMERA_CHN);
    
    return 0;
}

2.1. RV1126 模块初始化并启动 VI 工作

RV1126模块的初始化，包括VI模块、VENC模块的初始化，初始化上述模块后，则调用RK_MPI_VI_StartStream启动VI开始采集摄像头的视频流。关于VI模块、VENC模块的初始化参数这里就不阐述了，因为之前的课程里面已经讲了很多次。

2.2. rockx 人脸检测模块的初始化

复制代码 
    rockx_config_t * face_detect_config = rockx_create_config();
    rockx_add_config(face_detect_config, ROCKX_CONFIG_DATA_PATH, "/userdata/rockx_data");

    rockx_handle_t face_detect_handle;
    rockx_ret_t face_detect_ret;

    rockx_module_t face_detect_module = ROCKX_MODULE_FACE_DETECTION_V2;
    face_detect_ret = rockx_create(&face_detect_handle, face_detect_module, face_detect_config, 0);
    if(face_detect_ret != ROCKX_RET_SUCCESS)
    {
        printf("rockx_creat face_detect failed...\n");
        return -1;
    }
    rockx_release_config(face_detect_config);

这段代码是初始化rockx的模块，首先要使用rockx_create_config 分配rockx_config_t结构体，并使用rockx_add_config 把对应的rockx路径配置进去，在我们的板子里面在**/userdata/rockx_data** 里面,并使用rockx_create 创建rockx_handle_t句柄，rockx_create的传参第一个参数rockx_handle_t结构体指针、 第二个参数rockx_module_t是ROCKX_MODULE_FACE_DETECTION_V2 ， ROCKX_MODULE_FACE_DETECTION_V2是人脸检测的Version2模块、第三个参数是rockx_config_t结构体指针、第四个参数默认是0。

2.3. 使用 rockx 对 VI 模块的数据进行人脸检测处理

复制代码 
    rockx_image_t rv1126_rockx_image;
    memset(&rv1126_rockx_image, 0 ,sizeof(rv1126_rockx_image));
    rv1126_rockx_image.width = WIDTH;
    rv1126_rockx_image.height = HEIGHT;
    rv1126_rockx_image.pixel_format = ROCKX_PIXEL_FORMAT_YUV420SP_NV12;

    MEDIA_BUFFER mb;
    while(1)
    {
        mb = RK_MPI_SYS_GetMediaBuffer(RK_ID_VI, CAMERA_CHN, -1);
        if(!mb)
        {
            printf("Get Vi Stream break....\n");
            break;
        }

        rv1126_rockx_image.data = (uint8_t *)RK_MPI_MB_GetPtr(mb);
        rv1126_rockx_image.size = RK_MPI_MB_GetSize(mb);

        rockx_object_array_t face_detect_array;
        face_detect_ret = rockx_face_detect(face_detect_handle, &rv1126_rockx_image, &face_detect_array, NULL);
        if(face_detect_ret != ROCKX_RET_SUCCESS)
        {
            printf("face_detect failed....\n");
        }

这部分代码是整个DEMO的核心，也是ROCKX检测VI视频数据的核心。第一段是初始化rockx_image_t结构体，初始化需要传三个值分别是width = WIDTH(1920)、height = HEIGHT(1080)、pixel_format=ROCKX_PIXEL_FORMAT_YUV420SP_NV12。这三个值都需要和VI模块的配置是一样的。

初始化rockx_image_t 后，则需要通过RK_MPI_SYS_GetMediaBuffer获取每一帧VI模块的数据，并把每一帧VI模块的缓冲区和长度传输给rockx_image_t。具体的代码是rv1126_rockx_image.data = (uint8_t *)RK_MPI_MB_GetPtr(mb)(把每一帧VI缓冲区数据赋值到rockx_image_t的data)、rv1126_rockx_image.size = RK_MPI_MB_GetSize(mb)(把每一帧VI大小赋值到rockx_image_t的size)

赋值到rockx_image_t 后**，** 则调用rockx_face_detect 对每一帧的rockx_image_t图像进行人脸检测，并把人脸检测的结果输出到rockx_object_array_t 。 rockx_object_array_t的内容主要存储的是人脸检测数量和人脸检测区域信息(如：left、top、right、bottom的坐标信息)

2.4. 使用 opencv 对人脸检测的结果进行画框

复制代码 
        Mat rv1126_rockx_mat = Mat(HEIGHT, WIDTH, CV_8UC1, rv1126_rockx_image.data);

        for(int i = 0; i < face_detect_array.count; i++)
        {
            int left = face_detect_array.object[i].box.left;
            int top = face_detect_array.object[i].box.top;
            int w = face_detect_array.object[i].box.right - face_detect_array.object[i].box.left;
            int h = face_detect_array.object[i].box.bottom - face_detect_array.object[i].box.top;
            Rect boundingRect(left, top, w, h);
            rectangle(rv1126_rockx_mat, boundingRect, Scalar(255,255,0));
       }

检测完每一帧人脸数据后就需要对每个人脸区域进行画框了，这里画框是用opencv进行处理。首先要先创建OPENCV的Mat矩阵，Mat rv1126_image_mat = Mat(HEIGHT, WIDTH, CV_8UC1, rv1126_rockx_image.data) 。 创建完Mat之后，则需要根据rockx_object_array_t 的坐标信息进行画框，先循环遍历人脸的数量(rockx_object_array_t.count)，然后获取每一帧人脸的坐标信息,主要是left、top、right、bottom, 最后使用OPENCV的rectangle函数把坐标信息描绘出一个矩形表现出来。

2.5. 把处理后的数据发送到 VENC 模块

复制代码 
       RK_MPI_SYS_SendMediaBuffer(RK_ID_VENC, VENC_CHN, mb);
       RK_MPI_MB_ReleaseBuffer(mb);

把上述的数据处理完成之后则把每一帧数据传输给VENC模块，这里使用的API是RK_MPI_SYS_SendMediaBuffer **。**此时此刻VENC模块就有VENC码流数据了

2.6. 创建 rockx_face_detect_venc _thread 线程保存每一帧 H264 的编码码流数据

复制代码 
    pthread_t pid;
    pthread_create(&pid, NULL, rockx_face_detect_venc_thread, NULL);

void * rockx_face_detect_venc_thread(void * args)
{
    pthread_detach(pthread_self());
    FILE * face_detect_h264 = fopen("face_detect_venc.h264", "w+");

    MEDIA_BUFFER mb = NULL;

    while(1)
    {
        mb = RK_MPI_SYS_GetMediaBuffer(RK_ID_VENC, VENC_CHN, -1);
        if(!mb)
        {
            printf("Get Rockx_Venc Data berek...\n");
            break;
        }
        printf("Get Rockx_Venc Data Success...\n");

        fwrite(RK_MPI_MB_GetPtr(mb), RK_MPI_MB_GetSize(mb), 1, face_detect_h264);
        RK_MPI_MB_ReleaseBuffer(mb);
    }
    return NULL;
}

通过pthread_create创建venc码流线程，这个线程的名字是rockx_face_detect_venc thread。 在这个线程里面，通过RK_MPI_SYS_GetMediaBuffer 获取每一帧通过rockx人脸检测处理后的VENC码流数据，并用fwrite保存起来(fwrite(RK_MPI_MB_GetPtr(mb), RK_MPI_MB_GetSize(mb), 1, face_detect_h264))。

最终输出的结果是在视频中检测出对应的人脸并用opencv画矩形出来。

热门推荐
012026 年 AI 编程工具终极横评：Cursor vs Claude Code vs Copilot vs Windsurf02GitHub 镜像站点03【AI】2026 年具身智能模型和世界模型总结042026年6月AI大模型全景报告：GPT-5.6、Claude Opus 4.8、Gemini 3.5，中美AI三足鼎立谁主沉浮？052026 AI 编程工具终极实战指南：Cursor vs Claude Code vs Copilot，开发者该怎么选？06Claude Code、Codex、Cursor三分天下：2026年AI编程Agent生态全景剖析072026 年 AI 大模型 & AI 编程工具实战全总结082026年6月AI行业全景：从百模大战到Agent元年，这30天发生了什么？09AI科技热点日报 | 2026年07月01日10Trae国际版与国内版深度测评：AI原生IDE的双生花