基于RTMP的智慧数字人|AI数字人传输技术方案探讨
  rWdoMCxJ3E9e 2023年11月02日 37 0

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:

基于RTMP的智慧数字人|AI数字人传输技术方案探讨_智慧数字人

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。

【版权声明】本文内容来自摩杜云社区用户原创、第三方投稿、转载,内容版权归原作者所有。本网站的目的在于传递更多信息,不拥有版权,亦不承担相应法律责任。如果您发现本社区中有涉嫌抄袭的内容,欢迎发送邮件进行举报,并提供相关证据,一经查实,本社区将立刻删除涉嫌侵权内容,举报邮箱: cloudbbs@moduyun.com

  1. 分享:
最后一次编辑于 2023年11月08日 0

暂无评论

rWdoMCxJ3E9e