openh264 中背景检测功能源码分析

DogDaoDao2024-06-06 17:16

文件位置

  • openh264/codec/processing/src/BackgroundDetection.cpp

代码流程

  • 核心函数
    • 从代码流程可以看到实现背景检测的核心功能主要是CBackgroundDetection类中ForegroundBackgroundDivision函数和ForegroundDilationAndBackgroundErosion函数。

原理

  • 参数开关控制：(bool)bEnableBackgroundDetection

ForegroundBackgroundDivision 函数

  1. 功能：前景和背景粗略区分
  2. 原理过程：
  • 计算以宏块为单位的图像宽iPicWidthInOU和高iPicHeightInOU；
  • 计算每行的宏块个数iPicWidthInMb；
  • 声明并初始化指向Background OU数组的指针，该数组内部是宏块级OU 的数据信息；
  • 双层for 循环遍历所有宏块级 OU单元；
    • 调用GetOUParameters函数获取 OU 参数，主要是iSD、iSAD、ISD、iMAD、iMinSubMad、iMaxDiffSubSd；
    • 当前 OU背景标志iBackgroundFlag赋值 0；
    • 如果当前OU的平均绝对偏差（Mean Absolute Difference,MAD）大于63，则跳过当前OU；移动到下一个 OU；
    • 如果当前 OU 的iMaxDiffSubSd小于等于iSAD / 8或者iMaxDiffSubSd小于等于16x8，且iSAD小于2x16x16x2；
      • 如果iSAD小于16*8，则当前OU背景标志iBackgroundFlag 赋值 1；
      • 否则，当前OU背景标志iBackgroundFlag 的计算公式如下：
        • 如果当前 OU 的iSAD 小于2x16x16，判断当前 OU 的iSD是否小于iSADx3 / 4，将大小关系结果赋值给iBackgroundFlag；
        • 否则，判断当前 OU 的iSDx2 是否小于当前 OU 的iSAD，将大小关系结果赋值给iBackgroundFlag；
    • 移动到下一个宏块级OU；
  1. 原理图 ：
  2. 相关源码：
  • ForegroundBackgroundDivision函数
cpp 复制代码 
void CBackgroundDetection::ForegroundBackgroundDivision (vBGDParam* pBgdParam) {
  int32_t iPicWidthInOU         = pBgdParam->iBgdWidth  >> LOG2_BGD_OU_SIZE;
  int32_t iPicHeightInOU        = pBgdParam->iBgdHeight >> LOG2_BGD_OU_SIZE;
  int32_t iPicWidthInMb         = (15 + pBgdParam->iBgdWidth) >> 4;

  SBackgroundOU* pBackgroundOU = pBgdParam->pOU_array;

  for (int32_t j = 0; j < iPicHeightInOU; j ++) {
    for (int32_t i = 0; i < iPicWidthInOU; i++) {
      GetOUParameters (pBgdParam->pCalcRes, (j * iPicWidthInMb + i) << (LOG2_BGD_OU_SIZE - LOG2_MB_SIZE), iPicWidthInMb,
                       pBackgroundOU);

      pBackgroundOU->iBackgroundFlag = 0;
      if (pBackgroundOU->iMAD > 63) {
        pBackgroundOU++;
        continue;
      }
      if ((pBackgroundOU->iMaxDiffSubSd <= pBackgroundOU->iSAD >> 3
           || pBackgroundOU->iMaxDiffSubSd <= (BGD_OU_SIZE * Q_FACTOR))
          && pBackgroundOU->iSAD < (BGD_THD_SAD << 1)) { //BGD_OU_SIZE*BGD_OU_SIZE>>2
        if (pBackgroundOU->iSAD <= BGD_OU_SIZE * Q_FACTOR) {
          pBackgroundOU->iBackgroundFlag = 1;
        } else {
          pBackgroundOU->iBackgroundFlag = pBackgroundOU->iSAD < BGD_THD_SAD ?
                                           (pBackgroundOU->iSD < (pBackgroundOU->iSAD * 3) >> 2) :
                                           (pBackgroundOU->iSD << 1 < pBackgroundOU->iSAD);
        }
      }
      pBackgroundOU++;
    }
  }
}
  • GetOUParameters函数
cpp 复制代码 
void CBackgroundDetection::GetOUParameters (SVAACalcResult* sVaaCalcInfo, int32_t iMbIndex, int32_t iMbWidth,
    SBackgroundOU* pBgdOU) {
  int32_t       iSubSD[4];
  uint8_t       iSubMAD[4];
  int32_t       iSubSAD[4];

  uint8_t (*pMad8x8)[4];
  int32_t (*pSad8x8)[4];
  int32_t (*pSd8x8)[4];

  pSad8x8 = sVaaCalcInfo->pSad8x8;
  pMad8x8 = sVaaCalcInfo->pMad8x8;
  pSd8x8  = sVaaCalcInfo->pSumOfDiff8x8;

  iSubSAD[0] = pSad8x8[iMbIndex][0];
  iSubSAD[1] = pSad8x8[iMbIndex][1];
  iSubSAD[2] = pSad8x8[iMbIndex][2];
  iSubSAD[3] = pSad8x8[iMbIndex][3];

  iSubSD[0] = pSd8x8[iMbIndex][0];
  iSubSD[1] = pSd8x8[iMbIndex][1];
  iSubSD[2] = pSd8x8[iMbIndex][2];
  iSubSD[3] = pSd8x8[iMbIndex][3];

  iSubMAD[0] = pMad8x8[iMbIndex][0];
  iSubMAD[1] = pMad8x8[iMbIndex][1];
  iSubMAD[2] = pMad8x8[iMbIndex][2];
  iSubMAD[3] = pMad8x8[iMbIndex][3];

  pBgdOU->iSD   = iSubSD[0] + iSubSD[1] + iSubSD[2] + iSubSD[3];
  pBgdOU->iSAD  = iSubSAD[0] + iSubSAD[1] + iSubSAD[2] + iSubSAD[3];
  pBgdOU->iSD   = WELS_ABS (pBgdOU->iSD);

  // get the max absolute difference (MAD) of OU and min value of the MAD of sub-blocks of OU
  pBgdOU->iMAD = WELS_MAX (WELS_MAX (iSubMAD[0], iSubMAD[1]), WELS_MAX (iSubMAD[2], iSubMAD[3]));
  pBgdOU->iMinSubMad = WELS_MIN (WELS_MIN (iSubMAD[0], iSubMAD[1]), WELS_MIN (iSubMAD[2], iSubMAD[3]));

  // get difference between the max and min SD of the SDs of sub-blocks of OU
  pBgdOU->iMaxDiffSubSd = WELS_MAX (WELS_MAX (iSubSD[0], iSubSD[1]), WELS_MAX (iSubSD[2], iSubSD[3])) -
                          WELS_MIN (WELS_MIN (iSubSD[0], iSubSD[1]), WELS_MIN (iSubSD[2], iSubSD[3]));
}

ForegroundDilationAndBackgroundErosion 函数

  1. 功能：前景膨胀和背景腐蚀
  2. 函数关系图 ：
  3. 过程：
  • 定义iPicStrideUV变量存储 图像UV 分量步长；
  • 计算宏块级OU 的宽iPicWidthInOU和高iPicHeightInOU；
  • 计算宏块级OU 的 UV 步长iOUStrideUV；
  • 计算每行宏块个数iPicWidthInMb；
  • 定义了一个指向SBackgroundOU结构体的指针pBackgroundOU，用来访问背景检测单元数组；
  • 定义了一个指向int8_t的指针pVaaBackgroundMbFlag，用来访问背景宏块标志数组；
  • 定义了一个长度为4的指针数组pOUNeighbours，用来存储当前OU的四个邻居OU的指针；
  • 将pOUNeighbours[2]（即顶部OU）初始化为pBackgroundOU；
  • 外层 for 循环 遍历每一列宏块级OU；
    • 初始化pRowSkipFlag指针，指向当前行的宏块标志；
    • 初始化 pOUNeighbours[0] （左OU） 为pBackgroundOU；
    • 计算pOUNeighbours[3]（底OU）；
    • 内存 for 循环遍历每一行宏块级OU；
      • 计算pOUNeighbours[1]（即右OU）;
      • 如果当前 OU pBackgroundOU的iBackgroundFlag为真；
        • 调用ForegroundDilation函数进行前景膨胀；
      • 否则；
        • 调用BackgroundErosion函数进行背景腐蚀；
      • 调用UpperOUForegroundCheck函数检测上方 OU是否为前景；
      • 调用SetBackgroundMbFlag函数设置背景宏块标志；
      • 更新pRowSkipFlag，pOUNeighbours、pBackgroundOU为下一个 宏块级OU 做准备；
    • 更新pOUNeighbours[2]和pVaaBackgroundMbFlag指针，为下一行OU的处理做准备。
  1. 原理图 ：
  2. 相关源码：
  • ForegroundDilationAndBackgroundErosion函数
cpp 复制代码 
void CBackgroundDetection::ForegroundDilationAndBackgroundErosion (vBGDParam* pBgdParam) {
  int32_t iPicStrideUV          = pBgdParam->iStride[1];
  int32_t iPicWidthInOU         = pBgdParam->iBgdWidth  >> LOG2_BGD_OU_SIZE;
  int32_t iPicHeightInOU        = pBgdParam->iBgdHeight >> LOG2_BGD_OU_SIZE;
  int32_t iOUStrideUV           = iPicStrideUV << (LOG2_BGD_OU_SIZE - 1);
  int32_t iPicWidthInMb         = (15 + pBgdParam->iBgdWidth) >> 4;

  SBackgroundOU* pBackgroundOU = pBgdParam->pOU_array;
  int8_t*        pVaaBackgroundMbFlag = (int8_t*)pBgdParam->pBackgroundMbFlag;
  SBackgroundOU* pOUNeighbours[4];//0: left; 1: right; 2: top; 3: bottom

  pOUNeighbours[2]      = pBackgroundOU;//top OU
  for (int32_t j = 0; j < iPicHeightInOU; j ++) {
    int8_t* pRowSkipFlag = pVaaBackgroundMbFlag;
    pOUNeighbours[0]    = pBackgroundOU;//left OU
    pOUNeighbours[3]    = pBackgroundOU + (iPicWidthInOU & ((j == iPicHeightInOU - 1) - 1)); //bottom OU
    for (int32_t i = 0; i < iPicWidthInOU; i++) {
      pOUNeighbours[1] = pBackgroundOU + (i < iPicWidthInOU - 1); //right OU

      if (pBackgroundOU->iBackgroundFlag)
        ForegroundDilation (pBackgroundOU, pOUNeighbours, pBgdParam, j * iOUStrideUV + (i << LOG2_BGD_OU_SIZE_UV));
      else
        BackgroundErosion (pBackgroundOU, pOUNeighbours);

      // check the up OU
      if (j > 1 && i > 0 && i < iPicWidthInOU - 1 && pOUNeighbours[2]->iBackgroundFlag == 1) {
        UpperOUForegroundCheck (pOUNeighbours[2], pRowSkipFlag - OU_SIZE_IN_MB * iPicWidthInMb, iPicWidthInOU, iPicWidthInMb);
      }

      SetBackgroundMbFlag (pRowSkipFlag, iPicWidthInMb, pBackgroundOU->iBackgroundFlag);

      // preparation for the next OU
      pRowSkipFlag += OU_SIZE_IN_MB;
      pOUNeighbours[0] = pBackgroundOU;
      pOUNeighbours[2]++;
      pOUNeighbours[3]++;
      pBackgroundOU++;
    }
    pOUNeighbours[2]      = pBackgroundOU - iPicWidthInOU;
    pVaaBackgroundMbFlag += OU_SIZE_IN_MB * iPicWidthInMb;
  }
}
  • ForegroundDilation函数
cpp 复制代码 
inline void CBackgroundDetection::ForegroundDilation (SBackgroundOU* pBackgroundOU, SBackgroundOU* pOUNeighbours[],
    vBGDParam* pBgdParam, int32_t iChromaSampleStartPos) {
  int32_t iPicStrideUV = pBgdParam->iStride[1];
  int32_t iSumNeighBackgroundFlags = pOUNeighbours[0]->iBackgroundFlag + pOUNeighbours[1]->iBackgroundFlag +
                                      pOUNeighbours[2]->iBackgroundFlag + pOUNeighbours[3]->iBackgroundFlag;

  if (pBackgroundOU->iSAD > BGD_OU_SIZE * Q_FACTOR) {
    switch (iSumNeighBackgroundFlags) {
    case 0:
    case 1:
      pBackgroundOU->iBackgroundFlag = 0;
      break;
    case 2:
    case 3:
      pBackgroundOU->iBackgroundFlag = !ForegroundDilation23Luma (pBackgroundOU, pOUNeighbours);

      // chroma component check
      if (pBackgroundOU->iBackgroundFlag == 1) {
        int8_t iNeighbourForegroundFlags = (!pOUNeighbours[0]->iBackgroundFlag) | ((!pOUNeighbours[1]->iBackgroundFlag) << 1)
                                            | ((!pOUNeighbours[2]->iBackgroundFlag) << 2) | ((!pOUNeighbours[3]->iBackgroundFlag) << 3);
        pBackgroundOU->iBackgroundFlag = !ForegroundDilation23Chroma (iNeighbourForegroundFlags, iChromaSampleStartPos,
                                         iPicStrideUV, pBgdParam);
      }
      break;
    default:
      break;
    }
  }
}
  • BackgroundErosion函数
cpp 复制代码 
inline void CBackgroundDetection::BackgroundErosion (SBackgroundOU* pBackgroundOU, SBackgroundOU* pOUNeighbours[]) {
  if (pBackgroundOU->iMaxDiffSubSd <= (BGD_OU_SIZE * Q_FACTOR)) { //BGD_OU_SIZE*BGD_OU_SIZE>>2
    int32_t iSumNeighBackgroundFlags = pOUNeighbours[0]->iBackgroundFlag + pOUNeighbours[1]->iBackgroundFlag +
                                       pOUNeighbours[2]->iBackgroundFlag + pOUNeighbours[3]->iBackgroundFlag;
    int32_t sumNbrBGsad = (pOUNeighbours[0]->iSAD & (-pOUNeighbours[0]->iBackgroundFlag)) + (pOUNeighbours[2]->iSAD &
                          (-pOUNeighbours[2]->iBackgroundFlag))
                          + (pOUNeighbours[1]->iSAD & (-pOUNeighbours[1]->iBackgroundFlag)) + (pOUNeighbours[3]->iSAD &
                              (-pOUNeighbours[3]->iBackgroundFlag));
    if (pBackgroundOU->iSAD * iSumNeighBackgroundFlags <= (3 * sumNbrBGsad) >> 1) {
      if (iSumNeighBackgroundFlags == 4) {
        pBackgroundOU->iBackgroundFlag = 1;
      } else {
        if ((pOUNeighbours[0]->iBackgroundFlag & pOUNeighbours[1]->iBackgroundFlag)
            || (pOUNeighbours[2]->iBackgroundFlag & pOUNeighbours[3]->iBackgroundFlag)) {
          pBackgroundOU->iBackgroundFlag = !ForegroundDilation23Luma (pBackgroundOU, pOUNeighbours);
        }
      }
    }
  }
}
  • UpperOUForegroundCheck函数
cpp 复制代码 
inline void CBackgroundDetection::UpperOUForegroundCheck (SBackgroundOU* pCurOU, int8_t* pBackgroundMbFlag,
    int32_t iPicWidthInOU, int32_t iPicWidthInMb) {
  if (pCurOU->iSAD > BGD_OU_SIZE * Q_FACTOR) {
    SBackgroundOU* pOU_L = pCurOU - 1;
    SBackgroundOU* pOU_R = pCurOU + 1;
    SBackgroundOU* pOU_U = pCurOU - iPicWidthInOU;
    SBackgroundOU* pOU_D = pCurOU + iPicWidthInOU;
    if (pOU_L->iBackgroundFlag + pOU_R->iBackgroundFlag + pOU_U->iBackgroundFlag + pOU_D->iBackgroundFlag <= 1) {
      SetBackgroundMbFlag (pBackgroundMbFlag, iPicWidthInMb, 0);
      pCurOU->iBackgroundFlag = 0;
    }
  }
}
  • ForegroundDilation23Luma函数
cpp 复制代码 
inline bool CBackgroundDetection::ForegroundDilation23Luma (SBackgroundOU* pBackgroundOU,
    SBackgroundOU* pOUNeighbours[]) {
  SBackgroundOU* pOU_L = pOUNeighbours[0];
  SBackgroundOU* pOU_R = pOUNeighbours[1];
  SBackgroundOU* pOU_U = pOUNeighbours[2];
  SBackgroundOU* pOU_D = pOUNeighbours[3];

  if (pBackgroundOU->iMAD > pBackgroundOU->iMinSubMad << 1) {
    int32_t iMaxNbrForegroundMad;
    int32_t iMaxNbrBackgroundMad;
    int32_t aBackgroundMad[4];
    int32_t aForegroundMad[4];

    aForegroundMad[0] = (pOU_L->iBackgroundFlag - 1) & pOU_L->iMAD;
    aForegroundMad[1] = (pOU_R->iBackgroundFlag - 1) & pOU_R->iMAD;
    aForegroundMad[2] = (pOU_U->iBackgroundFlag - 1) & pOU_U->iMAD;
    aForegroundMad[3] = (pOU_D->iBackgroundFlag - 1) & pOU_D->iMAD;
    iMaxNbrForegroundMad = WELS_MAX (WELS_MAX (aForegroundMad[0], aForegroundMad[1]), WELS_MAX (aForegroundMad[2],
                                     aForegroundMad[3]));

    aBackgroundMad[0] = ((!pOU_L->iBackgroundFlag) - 1) & pOU_L->iMAD;
    aBackgroundMad[1] = ((!pOU_R->iBackgroundFlag) - 1) & pOU_R->iMAD;
    aBackgroundMad[2] = ((!pOU_U->iBackgroundFlag) - 1) & pOU_U->iMAD;
    aBackgroundMad[3] = ((!pOU_D->iBackgroundFlag) - 1) & pOU_D->iMAD;
    iMaxNbrBackgroundMad = WELS_MAX (WELS_MAX (aBackgroundMad[0], aBackgroundMad[1]), WELS_MAX (aBackgroundMad[2],
                                     aBackgroundMad[3]));

    return ((iMaxNbrForegroundMad > pBackgroundOU->iMinSubMad << 2) || (pBackgroundOU->iMAD > iMaxNbrBackgroundMad << 1
            && pBackgroundOU->iMAD <= (iMaxNbrForegroundMad * 3) >> 1));
  }
  return 0;
}
  • ForegroundDilation23Chroma函数
cpp 复制代码 
inline bool CBackgroundDetection::ForegroundDilation23Chroma (int8_t iNeighbourForegroundFlags,
    int32_t iStartSamplePos, int32_t iPicStrideUV, vBGDParam* pBgdParam) {
  static const int8_t kaOUPos[4]        = {OU_LEFT, OU_RIGHT, OU_TOP, OU_BOTTOM};
  int32_t       aEdgeOffset[4]          = {0, BGD_OU_SIZE_UV - 1, 0, iPicStrideUV* (BGD_OU_SIZE_UV - 1)};
  int32_t       iStride[4]              = {iPicStrideUV, iPicStrideUV, 1, 1};

  // V component first, high probability because V stands for red color and human skin colors have more weight on this component
  for (int32_t i = 0; i < 4; i++) {
    if (iNeighbourForegroundFlags & kaOUPos[i]) {
      uint8_t* pRefC = pBgdParam->pRef[2] + iStartSamplePos + aEdgeOffset[i];
      uint8_t* pCurC = pBgdParam->pCur[2] + iStartSamplePos + aEdgeOffset[i];
      if (CalculateAsdChromaEdge (pRefC, pCurC, iStride[i]) > BGD_THD_ASD_UV) {
        return 1;
      }
    }
  }
  // U component, which stands for blue color, low probability
  for (int32_t i = 0; i < 4; i++) {
    if (iNeighbourForegroundFlags & kaOUPos[i]) {
      uint8_t* pRefC = pBgdParam->pRef[1] + iStartSamplePos + aEdgeOffset[i];
      uint8_t* pCurC = pBgdParam->pCur[1] + iStartSamplePos + aEdgeOffset[i];
      if (CalculateAsdChromaEdge (pRefC, pCurC, iStride[i]) > BGD_THD_ASD_UV) {
        return 1;
      }
    }
  }

  return 0;
}
  • CalculateAsdChromaEdge函数
cpp 复制代码 
inline int32_t CBackgroundDetection::CalculateAsdChromaEdge (uint8_t* pOriRef, uint8_t* pOriCur, int32_t iStride) {
  int32_t ASD = 0;
  int32_t idx;
  for (idx = 0; idx < BGD_OU_SIZE_UV; idx++) {
    ASD += *pOriCur - *pOriRef;
    pOriRef += iStride;
    pOriCur += iStride;
  }
  return WELS_ABS (ASD);
}
相关推荐
兔云程序1 小时前
字节跳动收购Oladance耳机：强化音频技术，加速VR/AR生态布局
ar·音视频·vr
jndingxin2 小时前
OpenCV视频I/O(14）创建和写入视频文件的类:VideoWriter介绍
人工智能·opencv·音视频
芯橦7 小时前
【瑞昱RTL8763E】音频
单片机·嵌入式硬件·mcu·物联网·音视频·visual studio code·智能手表
9527华安8 小时前
FPGA实现PCIE视频采集转HDMI输出，基于XDMA中断架构，提供3套工程源码和技术支持
fpga开发·音视频·pcie·xdma·ov5640·hdmi
speop12 小时前
【笔记】I/O总结王道强化视频笔记
笔记·音视频
北城青14 小时前
WebRTC Connection Negotiate解决
运维·服务器·webrtc
cuijiecheng201817 小时前
音视频入门基础：FLV专题（13）——FFmpeg源码中，解析任意Type值的SCRIPTDATAVALUE类型的实现
ffmpeg·音视频
Envyᥫᩣ18 小时前
《ASP.NET Web Forms 实现视频点赞功能的完整示例》
前端·asp.net·音视频·视频点赞
applebomb19 小时前
【2024】uniapp 接入声网音频RTC【H5+Android】Unibest模板下Vue3+Typescript
typescript·uniapp·rtc·声网·unibest·agora
天天讯通19 小时前
网页WebRTC电话和软电话哪个好用？
webrtc
热门推荐
01【经验分享】Ubuntu22.04安装微信(linux官方版)02组基轨迹建模 GBTM的介绍与实现（Stata 或 R）032024年高教社杯数学建模国赛C题超详细解题思路分析04CTF网络安全大赛简单的web抓包题目：HEADache05安卓系列机型永久去除data分区加密 详细步骤解析06【2024数模国赛赛题思路公开】国赛B题思路丨附可运行代码丨无偿自提07pve(Proxmox VE)安装i225v网卡驱动08Ubuntu24.04安装中文输入法09【2024高教社杯全国大学生数学建模竞赛】B题 生产过程中的决策问题——解题思路 代码 论文10分享几个惊艳的 Three.js 个人站点