基于RTMP的智慧数字人|AI数字人传输技术方案探讨

技术背景

随着智慧数字人、AI数字人的兴起，越来越多的公司着手构建​全息、真实感数字角色等技术合成的数字仿真人虚拟形象，通过“虚拟形象+语音交互（T-T-S、ASR）+自然语言理解（NLU）+深度学习”，构建适用于数字客服、虚拟展厅讲解、 智慧城市、智慧医疗、智慧教育等场景，通过人机可视化语音交互，释放人员基础劳动力，降低运营成本，提升智慧交互体验。​

​一个有“温度”的智慧数字人，有多个维度组成，如图像识别、语音识别、语义理解等，本文主要阐述的是如何把这样一个智慧数字人，通过编码传输，以更低的延迟和好的体验，呈现给用户。​

​技术实现​

​本文以Windows平台为例，从技术角度探讨智慧数字人的实时编码传输。先上图：​

左侧是Unity采集、获取video Texture和AudioClip数据，编码打包后，然后通过RTMP推送到服务端，右下侧实时拉取RTMP流数据播放，整体延迟在毫秒级。

视频采集这块，实现了Unity获取到的Texture数据的采集、摄像头采集、屏幕采集三大类：

public void SelVideoPushType(int type)
    {
        switch (type)
        {
            case 0:
                video_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_VIDEO_OPTION.NT_PB_E_VIDEO_OPTION_LAYER;    //采集Unity窗体
                break;
            case 1:
                video_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_VIDEO_OPTION.NT_PB_E_VIDEO_OPTION_CAMERA;   //采集摄像头
                break;
            case 2:
                video_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_VIDEO_OPTION.NT_PB_E_VIDEO_OPTION_SCREEN;   //采集屏幕
                break;
            case 3:
                video_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_VIDEO_OPTION.NT_PB_E_VIDEO_OPTION_NO_VIDEO; //不采集视频
                break;
        }

        Debug.Log("SelVideoPushType type: " + type + " video_push_type: " + video_push_type_);
    }

音频采集部分，我们主要实现了采集AudioClip的声音、麦克风、扬声器、还有两路AudioClip的音频混音：

public void SelAudioPushType(int type)
    {
        switch (type)
        {
            case 0:
                audio_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_AUDIO_OPTION.NT_PB_E_AUDIO_OPTION_EXTERNAL_PCM_DATA;    //采集Unity声音
                break;
            case 1:
                audio_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_AUDIO_OPTION.NT_PB_E_AUDIO_OPTION_CAPTURE_MIC;  //采集麦克风
                break;
            case 2:
                audio_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_AUDIO_OPTION.NT_PB_E_AUDIO_OPTION_CAPTURE_SPEAKER;  //采集扬声器
                break;
            case 3:
                audio_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_AUDIO_OPTION.NT_PB_E_AUDIO_OPTION_TWO_EXTERNAL_PCM_MIXER;  //两路Unity AudioClip混音
                break;
            case 4:
                audio_push_type_ = (uint)NTSmartPublisherDefine.NT_PB_E_AUDIO_OPTION.NT_PB_E_AUDIO_OPTION_NO_AUDIO;   //不采集音频
                break;
        }

        Debug.Log("SelAudioPushType type: " + type + " audio_push_type: " + audio_push_type_);
    }

为了便于测试延迟，在页面加了个简单的时间日期刷新：

//获取当前时间
        GameObject.Find("Canvas/Panel/LableText").GetComponent<Text>().text = string.Format("{0:D2}:{1:D2}:{2:D2}:{3:D2} " + "{4:D4}/{5:D2}/{6:D2}",
            DateTime.Now.Hour, DateTime.Now.Minute, DateTime.Now.Second, DateTime.Now.Millisecond,
            DateTime.Now.Year, DateTime.Now.Month, DateTime.Now.Day);

Unity窗体或Camera采集，可以从Texuture拿到数据，从而获取到rgb数据，投递到封装的wrapper层，实现编码传输。

if (texture_ == null || video_width_ != Screen.width || video_height_ != Screen.height)
  {
      Debug.Log("OnPostRender screen changed++ scr_width: " + Screen.width + " scr_height: " + Screen.height);

      if (screen_image_ != IntPtr.Zero)
      {
          Marshal.FreeHGlobal(screen_image_);
          screen_image_ = IntPtr.Zero;
      }

      if (texture_ != null)
      {
          UnityEngine.Object.Destroy(texture_);
          texture_ = null;
      }

      video_width_ = Screen.width;
      video_height_ = Screen.height;

      texture_ = new Texture2D(video_width_, video_height_, TextureFormat.BGRA32, false);

      screen_image_ = Marshal.AllocHGlobal(video_width_ * 4 * video_height_);

      Debug.Log("OnPostRender screen changed--");

      return;
  }

  texture_.ReadPixels(new Rect(0, 0, video_width_, video_height_), 0, 0, false);
  texture_.Apply();

摄像头和屏幕采集，可以直接在封装层实现，如果需要做预览，只需要把数据回到Unity，通过RawImage实时刷新Texture显示即可。

通过封装层实现数据预览：

public bool StartPreview()
    {
        if(CheckPublisherHandleAvailable() == false)
            return false;

        video_preview_image_callback_ = new NT_PB_SDKVideoPreviewImageCallBack(SDKVideoPreviewImageCallBack);

        NTSmartPublisherSDK.NT_PB_SetVideoPreviewImageCallBack(publisher_handle_, (int)NTSmartPublisherDefine.NT_PB_E_IMAGE_FORMAT.NT_PB_E_IMAGE_FORMAT_RGB32, IntPtr.Zero, video_preview_image_callback_);

        if (NTBaseCodeDefine.NT_ERC_OK != NTSmartPublisherSDK.NT_PB_StartPreview(publisher_handle_, 0, IntPtr.Zero))
        {
            if (0 == publisher_handle_count_)
            {
                NTSmartPublisherSDK.NT_PB_Close(publisher_handle_);
                publisher_handle_ = IntPtr.Zero;
            }

            return false;
        }

        publisher_handle_count_++;

        is_previewing_ = true;

        return true;
    }

    public void StopPreview()
    {
        if (is_previewing_ == false) return;

        is_previewing_ = false;

        publisher_handle_count_--;
        NTSmartPublisherSDK.NT_PB_StopPreview(publisher_handle_);

        if (0 == publisher_handle_count_)
        {
            NTSmartPublisherSDK.NT_PB_Close(publisher_handle_);
            publisher_handle_ = IntPtr.Zero;
        }
    }

预览数据回调：

//预览数据回调
  public void SDKVideoPreviewImageCallBack(IntPtr handle, IntPtr user_data, IntPtr image)
  {
      NT_PB_Image pb_image = (NT_PB_Image)Marshal.PtrToStructure(image, typeof(NT_PB_Image));

      NT_VideoFrame pVideoFrame = new NT_VideoFrame();

      pVideoFrame.width_ = pb_image.width_;
      pVideoFrame.height_ = pb_image.height_;

      pVideoFrame.stride_ = pb_image.stride_[0];

      Int32 argb_size = pb_image.stride_[0] * pb_image.height_;

      pVideoFrame.plane_data_ = new byte[argb_size];

      if (argb_size > 0)
      {
          Marshal.Copy(pb_image.plane_[0],pVideoFrame.plane_data_,0, argb_size);
      }

      {
          cur_image_ = pVideoFrame;
      }
  }

音频采集这块，Unity环境下，主要是采集Unity的AudioClip数据，这块需要注意的是，PCM数据发送间隔，每隔10毫秒发一次，因为AudioClip的size比如可能只有十几秒或者几分钟，需要考虑的是，AudioClip数据采集播放完毕后，是loop的形式反复播放，还是静音帧的形式，只传视频，不传音频。

var pcm_data = new PCMData();
pcm_data.sample_rate_ = audio_clip_info_.audio_clip_.frequency;
pcm_data.channels_ = audio_clip_info_.audio_clip_.channels;
pcm_data.per_channel_sample_number_ = pcm_data.sample_rate_ / 100;

var pcm_sample = new float[pcm_data.sample_rate_ * pcm_data.channels_ / 100];

audio_clip_info_.audio_clip_.GetData(pcm_sample, audio_clip_info_.audio_clip_offset_);

var sample_length = sizeof(float) * pcm_sample.Length;

pcm_data.data_ = Marshal.AllocHGlobal(sample_length);
Marshal.Copy(pcm_sample, 0, pcm_data.data_, pcm_sample.Length);
pcm_data.size_ = (uint)sample_length;

publisher_wrapper_.OnPostAudioPCMFloatData(pcm_data.data_,
    pcm_data.size_,
    pcm_time_stamp_,
    pcm_data.sample_rate_,
    pcm_data.channels_,
    pcm_data.per_channel_sample_number_);

Marshal.FreeHGlobal(pcm_data.data_);
pcm_data.data_ = IntPtr.Zero;
pcm_data = null;

pcm_time_stamp_ += 10;  //时间戳自增10毫秒

如果要两路混音，只要再从Resources下面，获取另一路AudioClip数据，然后投递即可：

audio_clip_info_mix_ = new AudioClipInfo();
audio_clip_info_mix_.audio_clip_ = Resources.Load("AudioData/music") as AudioClip;

​数据投递，用以下接口：​

publisher_wrapper_.OnPostAudioExternalPCMFloatMixerData(pcm_data_mix.data_,
      pcm_data_mix.size_,
      pcm_time_stamp_mix_,
      pcm_data_mix.sample_rate_,
      pcm_data_mix.channels_,
      pcm_data_mix.per_channel_sample_number_);

数据采集投递过来后，我们以图层的形式投递过来，设置音视频编码参数，底层实现音视频编码：

/*
 * nt_publisher_wrapper.cs
 * nt_publisher_wrapper
 * 
 * Github: https://github.com/daniulive/SmarterStreaming
 * 
 * Created by DaniuLive on 2017/11/14.
 */  
private void SetCommonOptionToPublisherSDK()
  {
      if (!IsPublisherHandleAvailable())
      {
          Debug.Log("SetCommonOptionToPublisherSDK, publisher handle with null..");
          return;
      }

      NTSmartPublisherSDK.NT_PB_ClearLayersConfig(publisher_handle_, 0,
                      0, IntPtr.Zero);

      if (video_option_ == (uint)NTSmartPublisherDefine.NT_PB_E_VIDEO_OPTION.NT_PB_E_VIDEO_OPTION_LAYER)
      {
          // 第0层填充RGBA矩形, 目的是保证帧率, 颜色就填充全黑
          int red = 0;
          int green = 0;
          int blue = 0;
          int alpha = 255;

          NT_PB_RGBARectangleLayerConfig rgba_layer_c0 = new NT_PB_RGBARectangleLayerConfig();

          rgba_layer_c0.base_.type_ = (Int32)NTSmartPublisherDefine.NT_PB_E_LAYER_TYPE.NT_PB_E_LAYER_TYPE_RGBA_RECTANGLE;
          rgba_layer_c0.base_.index_ = 0;
          rgba_layer_c0.base_.enable_ = 1;
          rgba_layer_c0.base_.region_.x_ = 0;
          rgba_layer_c0.base_.region_.y_ = 0;
          rgba_layer_c0.base_.region_.width_ = video_width_;
          rgba_layer_c0.base_.region_.height_ = video_height_;

          rgba_layer_c0.base_.offset_ = Marshal.OffsetOf(rgba_layer_c0.GetType(), "base_").ToInt32();
          rgba_layer_c0.base_.cb_size_ = (uint)Marshal.SizeOf(rgba_layer_c0);

          rgba_layer_c0.red_ = System.BitConverter.GetBytes(red)[0];
          rgba_layer_c0.green_ = System.BitConverter.GetBytes(green)[0];
          rgba_layer_c0.blue_ = System.BitConverter.GetBytes(blue)[0];
          rgba_layer_c0.alpha_ = System.BitConverter.GetBytes(alpha)[0];

          IntPtr rgba_conf = Marshal.AllocHGlobal(Marshal.SizeOf(rgba_layer_c0));

          Marshal.StructureToPtr(rgba_layer_c0, rgba_conf, true);

          UInt32 rgba_r = NTSmartPublisherSDK.NT_PB_AddLayerConfig(publisher_handle_, 0,
                          rgba_conf, (int)NTSmartPublisherDefine.NT_PB_E_LAYER_TYPE.NT_PB_E_LAYER_TYPE_RGBA_RECTANGLE,
                          0, IntPtr.Zero);

          Marshal.FreeHGlobal(rgba_conf);

          NT_PB_ExternalVideoFrameLayerConfig external_layer_c1 = new NT_PB_ExternalVideoFrameLayerConfig();

          external_layer_c1.base_.type_ = (Int32)NTSmartPublisherDefine.NT_PB_E_LAYER_TYPE.NT_PB_E_LAYER_TYPE_EXTERNAL_VIDEO_FRAME;
          external_layer_c1.base_.index_ = 1;
          external_layer_c1.base_.enable_ = 1;
          external_layer_c1.base_.region_.x_ = 0;
          external_layer_c1.base_.region_.y_ = 0;
          external_layer_c1.base_.region_.width_ = video_width_;
          external_layer_c1.base_.region_.height_ = video_height_;

          external_layer_c1.base_.offset_ = Marshal.OffsetOf(external_layer_c1.GetType(), "base_").ToInt32();
          external_layer_c1.base_.cb_size_ = (uint)Marshal.SizeOf(external_layer_c1);

          IntPtr external_layer_conf = Marshal.AllocHGlobal(Marshal.SizeOf(external_layer_c1));

          Marshal.StructureToPtr(external_layer_c1, external_layer_conf, true);

          UInt32 external_r = NTSmartPublisherSDK.NT_PB_AddLayerConfig(publisher_handle_, 0,
                          external_layer_conf, (int)NTSmartPublisherDefine.NT_PB_E_LAYER_TYPE.NT_PB_E_LAYER_TYPE_EXTERNAL_VIDEO_FRAME,
                          0, IntPtr.Zero);

          Marshal.FreeHGlobal(external_layer_conf);

      }
      else if (video_option_ == (uint)NTSmartPublisherDefine.NT_PB_E_VIDEO_OPTION.NT_PB_E_VIDEO_OPTION_CAMERA)
      {
          CameraInfo camera = cameras_[cur_sel_camera_index_];
          NT_PB_VideoCaptureCapability cap = camera.capabilities_[cur_sel_camera_resolutions_index_];

          SetVideoCaptureDeviceBaseParameter(camera.id_.ToString(), (UInt32)cap.width_, (UInt32)cap.height_);
      }

      SetFrameRate((uint)video_fps_);

      Int32 type = 0;   //软编码
      Int32 encoder_id = 1;
      UInt32 codec_id = (UInt32)NTCommonMediaDefine.NT_MEDIA_CODEC_ID.NT_MEDIA_CODEC_ID_H264;
      Int32 param1 = 0;

      SetVideoEncoder(type, encoder_id, codec_id, param1);

      SetVideoQualityV2(CalVideoQuality(video_width_, video_height_, is_h264_encoder_));

      SetVideoBitRate(CalBitRate(video_fps_, video_width_, video_height_));

      SetVideoMaxBitRate((CalMaxKBitRate(video_fps_, video_width_, video_height_, false)));

      SetVideoKeyFrameInterval((key_frame_interval_));

      if (is_h264_encoder_)
      {
          SetVideoEncoderProfile(1);
      }

      SetVideoEncoderSpeed(CalVideoEncoderSpeed(video_width_, video_height_, is_h264_encoder_));

      // 音频相关设置
      SetAuidoInputDeviceId(0);
      SetPublisherAudioCodecType(1);
      SetPublisherMute(is_mute_);
      SetEchoCancellation(0, 0);
      SetNoiseSuppression(0);
      SetAGC(0);
      SetVAD(0);
      SetInputAudioVolume(Convert.ToSingle(audio_input_volume_));
  }

编码打包后，可以调用推送接口，把打包后的数据，实时传到RTMP服务端：

public bool StartPublisher(String url)
    {
        if (CheckPublisherHandleAvailable() == false) return false;

        if (publisher_handle_ == IntPtr.Zero)
        {
            return false;
        }
        if (!String.IsNullOrEmpty(url))
        {
            NTSmartPublisherSDK.NT_PB_SetURL(publisher_handle_, url, IntPtr.Zero);
        }

        if (NTBaseCodeDefine.NT_ERC_OK != NTSmartPublisherSDK.NT_PB_StartPublisher(publisher_handle_, IntPtr.Zero))
        {
            if (0 == publisher_handle_count_)
            {
                NTSmartPublisherSDK.NT_PB_Close(publisher_handle_);
                publisher_handle_ = IntPtr.Zero;
            }

            is_publishing_ = false;

            return false;
        }

        publisher_handle_count_++;

        is_publishing_ = true;

        return true;
    }

    public void StopPublisher()
    {
        if (is_publishing_ == false) return;

        publisher_handle_count_--;
        NTSmartPublisherSDK.NT_PB_StopPublisher(publisher_handle_);

        if (0 == publisher_handle_count_)
        {
            NTSmartPublisherSDK.NT_PB_Close(publisher_handle_);
            publisher_handle_ = IntPtr.Zero;
        }

        is_publishing_ = false;
    }

RTMP传输这块，需要把Event状态回调给Unity，确保Unity实时处理网络异常：

Unity层处理：

public event Action<uint,string> OnLogEventMsg;

publisher_wrapper_.OnLogEventMsg += OnLogHandle;

private void OnLogHandle(uint arg1, string arg2)
{
    Debug.Log(arg2);
}

wrapper层处理：

private void PbEventCallBack(IntPtr handle, IntPtr user_data, 
      UInt32 event_id,
      Int64 param1,
      Int64 param2,
      UInt64 param3,
      UInt64 param4,
      [MarshalAs(UnmanagedType.LPStr)] String param5,
      [MarshalAs(UnmanagedType.LPStr)] String param6,
      IntPtr param7)
  {
      String event_log = "";

      switch (event_id)
      {
          case (uint)NTSmartPublisherDefine.NT_PB_E_EVENT_ID.NT_PB_E_EVENT_ID_CONNECTING:
              event_log = "连接中";
              if (!String.IsNullOrEmpty(param5))
              {
                  event_log = event_log + " url:" + param5;
              }
              break;

          case (uint)NTSmartPublisherDefine.NT_PB_E_EVENT_ID.NT_PB_E_EVENT_ID_CONNECTION_FAILED:
              event_log = "连接失败";
              if (!String.IsNullOrEmpty(param5))
              {
                  event_log = event_log + " url:" + param5;
              }
              break;

          case (uint)NTSmartPublisherDefine.NT_PB_E_EVENT_ID.NT_PB_E_EVENT_ID_CONNECTED:
              event_log = "已连接";
              if (!String.IsNullOrEmpty(param5))
              {
                  event_log = event_log + " url:" + param5;
              }
              break;

          case (uint)NTSmartPublisherDefine.NT_PB_E_EVENT_ID.NT_PB_E_EVENT_ID_DISCONNECTED:
              event_log = "断开连接";
              if (!String.IsNullOrEmpty(param5))
              {
                  event_log = event_log + " url:" + param5;
              }
              break;

          default:
              break;
      }

      if(OnLogEventMsg != null) OnLogEventMsg.Invoke(event_id, event_log);
  }

总结

以上是大概的流程，通过采集Unity的音视频数据，编码打包传输，发送到RTMP服务端，客户端直接拉取RTMP流数据，延迟在毫秒级，用户体验良好，在智慧数字人等交互场景，体验极佳。