使用FFmpeg进行视频会议录制后处理——转码与优化教程

ffmpeg -hide_banner -encoders | grep nvenc
# 输出包含 h264_nvenc / hevc_nvenc 即可用

# 输入：Zoom 原始 .zoom / .mp4  输出：标准 MP4，关键帧固定 2s，不重新编码视频
ffmpeg -y -i input.zoom 
  -c:v copy -c:a aac -b:a 96k 
  -movflags +faststart 
  -g 50 -keyint_min 50 -sc_threshold 0 
  output_meeting_pre.mp4

ffmpeg -y -i output_meeting_pre.mp4 
  -c:v libx264 -preset medium -crf 24 
  -profile:v high -level 4.1 
  -vf "scale='if(gt(iw,ih),1920,-2)':'if(gt(iw,ih),-2,1080)':force_original_aspect_ratio=decrease" 
  -c:a libfdk_aac -b:a 96k -ac 1 
  -movflags +faststart 
  meeting_archive_1080p.mp4

ffmpeg -y -hwaccel cuda -hwaccel_output_format cuda -i input.mp4 
  -c:v hevc_nvenc -preset p5 -rc vbr -cq 28 -b:v 0 
  -g 50 -keyint_min 50 
  -vf "scale_cuda=1920:1080:force_original_aspect_ratio=decrease" 
  -c:a aac -b:a 96k 
  meeting_hevc_gpu.mp4

# 生成 1080p / 720p / 480p 三码率 HLS，统一关键帧对齐
ffmpeg -y -i meeting_archive_1080p.mp4 
  -filter_complex 
  "[v:0]split=3[v1][v2][v3]; 
   [v1]scale=1920:1080:force_original_aspect_ratio=decrease[v1out]; 
   [v2]scale=1280:720:force_original_aspect_ratio=decrease[v2out]; 
   [v3]scale=854:480:force_original_aspect_ratio=decrease[v3out]" 
  -map "[v1out]" -c:v:0 libx264 -b:v:0 3500k -maxrate:v:0 3850k -bufsize:v:0 7000k 
  -map "[v2out]" -c:v:1 libx264 -b:v:1 2000k -maxrate:v:1 2200k -bufsize:v:1 4000k 
  -map "[v3out]" -c:v:2 libx264 -b:v:2 1000k -maxrate:v:2 1100k -bufsize:v:2 2000k 
  -map a:0 -c:a aac -b:a 96k -ac 1 
  -f hls -hls_time 6 -hls_playlist_type vod -hls_flags independent_segments 
  -master_pl_name master.m3u8 
  -hls_segment_filename "stream_%v/data%03d.ts" 
  stream_%v.m3u8

# 检查关键帧间隔是否符合预期
ffprobe -select_streams v -show_entries frame=key_frame,pkt_pts_time -of csv meeting_archive_1080p.mp4 | awk -F, '$2==1{print $3}' | awk 'NR>1{print $1-prev}{prev=$1}' | sort -n | tail -5
# 若输出最大值 > 3 秒，需复核 -g 参数生效情况

graph LR
A[会议平台 Webhook] --> B(对象存储触发事件)
B --> C{消息队列 Kafka/RabbitMQ}
C --> D[Worker 节点拉取任务]
D --> E[FFmpeg 转码脚本 (参数化模板)]
E --> F[ffprobe 质检通过?]
F -- 是 --> G[写入归档 Bucket + 更新数据库]
F -- 否 --> H[告警 + 进入人工复核队列]
G --> I[CDN 刷新预热]
I --> J[前端播放器加载 master.m3u8]

# 1. 查看完整流信息
ffprobe -v error -show_entries stream=codec_name,width,height,r_frame_rate,bitrate,duration -of default=nw=1 input.mp4

# 2. 计算 VMAF (需编译 libvmaf，对比原始与压缩后画质)
ffmpeg -i distorted.mp4 -i reference.mp4 -lavfi libvmaf -f null -

# 3. 批量重命名并按日期归档 (配合 find 使用)
find . -name "*.zoom" -exec sh -c 'ffmpeg -i "$1" -c copy "${1%.zoom}.mp4"' _ {} ;

# 语谱图可视化调试：ffmpeg -i in.wav -lavfi showspectrumpic=s=1280x720:legend=1:fscale=log out.png
ffmpeg -y -i input.mp4 
  -filter_complex 
  "[0:a] 
   highpass=f=80, 
   lowpass=f=8000, 
   afftdn=nr=12:nf=-25:tn=1, 
   arnndn=m='rnnoise-models/somnolent-hog.rnnn', 
   dynaudnorm=p=0.9:m=10:s=5:g=15, 
   loudnorm=I=-16:TP=-1.5:LRA=11:print_format=json 
  [aout]" 
  -map 0:v -map "[aout]" 
  -c:v copy -c:a libfdk_aac -b:a 96k 
  output_audio_enhanced.mp4

# 1. 生成静音段时间戳（JSON 供下游解析）
ffmpeg -i input.mp4 -af silencedetect=n=-40dB:d=1.5 -f null - 2>&1 | grep silencedetect

# 2. 一键生成“去静音/紧凑版”（在线培训/精华剪辑场景）
ffmpeg -y -i input.mp4 
  -vf "select='not(gt(volume,-40))',setpts=N/FRAME_RATE/TB" 
  -af "aselect='not(gt(volume,-40))',asetpts=N/SR/TB" 
  output_compact.mp4

# 核心思路：检测画面变化，仅在内容变化时输出帧，并写入正确 PTS
ffmpeg -y -i input.mp4 
  -vf "mpdecimate=hi=64*12:lo=64*8:frac=0.33,setpts=N/FRAME_RATE/TB" 
  -vsync vfr 
  -c:v libx264 -preset medium -crf 24 -profile:v high -level 4.1 
  -c:a copy 
  output_vfr.mp4

# 简易版：固定区域裁剪（假设共享区固定 1920x1080 左侧，画中画右侧 320x180）
ffmpeg -y -i input.mp4 
  -filter_complex 
  "[0:v]crop=1920:1080:0:0[share]; 
   [0:v]crop=320:180:1600:900[pip]; 
   [share]scale=1920:1080[share_up]; 
   [pip]scale=320:180[pip_up]" 
  -map "[share_up]" -c:v:0 libx264 -crf 20 -b:v:0 4000k 
  -map "[pip_up]" -c:v:1 libx264 -crf 28 -b:v:1 300k 
  -map 0:a -c:a copy 
  output_roi.mp4

# 需 FFmpeg 编译 --enable-libplacebo --enable-libopencv --enable-dnn
# 模型下载：opencv_face_detector_uint8.pb + opencv_face_detector.pbtxt
ffmpeg -y -i input.mp4 
  -vf "dnn_backend=opencv:model=opencv_face_detector_uint8.pb:config=opencv_face_detector.pbtxt:input=blob:output=detections, 
       gblur=sigma=30:steps=3:planes=0:1:2" 
  -c:v h264_nvenc -preset p4 -cq 24 
  -c:a copy 
  output_blur.mp4

# 伪代码：管线中插入 Python 步骤
# 1. ffmpeg 抽帧 1fps -> frames/
# 2. PaddleOCR / Tesseract 识别文本坐标
# 3. 正则匹配 身份证/手机/银行卡/密码 -> 生成 drawbox 滤镜脚本
# 4. ffmpeg 二次跑流应用遮盖

# 生成的动态遮盖脚本示例 (overlay_filter.txt)
# 格式：sendcmd 'ftime=12.5 drawbox=x=400:y=300:w=200:h=40:color=black@1:t=fill'

# ffmpeg-transcode-job.yaml
apiVersion: batch/v1
kind: Job
metadata:
  name: ffmpeg-transcode-{{MEETING_ID}}
  labels:
    app: ffmpeg-worker
    priority: "{{PRIORITY_CLASS}}"  # high/normal/low
spec:
  backoffLimit: 2
  activeDeadlineSeconds: 7200  # 2h 超时自动杀掉
  template:
    spec:
      priorityClassName: "{{PRIORITY_CLASS}}"
      restartPolicy: OnFailure
      volumes:
      - name: workdir
        emptyDir:
          sizeLimit: 50Gi      # 限制临时盘，防止爆盘
          medium: Memory       # 高优先级任务用内存盘加速 I/O
      - name: model-cache
        persistentVolumeClaim:
          claimName: rnnoise-models-pvc
      containers:
      - name: ffmpeg
        image: registry.internal/ffmpeg:7.0-cuda12-rnnoise
        resources:
          requests:
            nvidia.com/gpu: 1
            cpu: "4"
            memory: "16Gi"
            ephemeral-storage: "20Gi"
          limits:
            nvidia.com/gpu: 1
            cpu: "8"
            memory: "32Gi"
        env:
        - name: NVIDIA_VISIBLE_DEVICES
          value: all
        - name: FFMPEG_THREADS
          value: "4"           # 限制线程数，防止抢占同节点其他 Pod
        volumeMounts:
        - name: workdir
          mountPath: /workspace
        - name: model-cache
          mountPath: /models
        command: ["/bin/bash", "-c"]
        args:
        - |
          set -ex
          cd /workspace
          # 1. 下载源文件 (预签名 URL)
          aria2c -x 16 -s 16 "$SRC_URL" -o input.raw
          # 2. 执行转码 (参数化模板)
          /scripts/transcode_pipeline.sh input.raw output.mp4
          # 3. 质检
          /scripts/qc_check.sh output.mp4
          # 4. 上传归档
          coscli cp output.mp4 cos://archive-bucket/{{TENANT_ID}}/{{DATE}}/{{MEETING_ID}}.mp4
          # 5. 回调业务系统
          curl -X POST "$CALLBACK_URL" -d '{"status":"success","meeting_id":"{{MEETING_ID}}"}'

# 管线末端并行跑：音频抽取 -> Whisper.cpp / FunASR -> WebVTT
ffmpeg -i meeting.mp4 -vn -ac 1 -ar 16000 -f wav - | 
  whisper.cpp -m models/ggml-medium.bin -l zh -oj -f - > meeting.json
# 转 WebVTT (前端播放器原生支持字幕轨)
python3 json2vtt.py meeting.json > meeting.vtt

# 每 30 秒抽 1 帧 -> 批量送入 Qwen-VL / GPT-4o / LLaVA
# Prompt: "描述画面内容，判断是否为 PPT/代码/白板/人脸，提取标题关键词"
# 输出: [{"ts": 1200, "label": "PPT", "title": "Q3 预算执行情况", "ocr_text": "..."}]

# 结合 ASR 语义边界 + 视觉场景切换 -> 生成章节点
# FFmpeg 一键切片（无损关键帧切割）
ffmpeg -i meeting.mp4 -f segment -segment_times 300,1200,2400 -c copy -reset_timestamps 1 chapter_%03d.mp4

{
  "ts": "2024-05-20T10:00:00.123Z",
  "level": "INFO",
  "task_id": "transcode_abc123",
  "stage": "video_encode",
  "codec": "hevc_nvenc",
  "params": {"cq": 28, "preset": "p5", "gop": 50},
  "metrics": {"fps": 245.3, "vmaf": 94.2, "vram_mb": 1850, "duration_sec": 42.1},
  "output": {"size_mb": 312, "duration_sec": 3600}
}

# 1. 生成雪碧图 (每 10 秒 1 帧，10x10 拼图)
ffmpeg -i input.mp4 -vf "fps=1/10,scale=160:90,tile=10x10" -q:v 2 sprite.jpg
# 配合 WebVTT: 00:00:00.000 --> 00:00:10.000 {xywh=0,0,160,90}

# 2. 提取关键帧为图片 (场景检测)
ffmpeg -i input.mp4 -vf "select='gt(scene,0.4)',scale=320:180" -vsync vfr keyframe_%04d.jpg

# 3. 无损切割 (关键帧级，毫秒级精度)
ffmpeg -ss 00:10:00 -i input.mp4 -t 00:05:00 -c copy -avoid_negative_ts make_zero clip.mp4

# 4. 实时监控编码进度 (管线集成)
ffmpeg -i input.mp4 ... -progress pipe:1 -nostats 2>/dev/null | grep -E 'out_time_ms|progress=end'

# 5. 容器健康检查 (K8s livenessProbe)
ffmpeg -v error -i /workspace/input.mp4 -f null - || exit 1