OpenCV，ffmpeg与libx264进行H.264编码（CPU与GPU版）
CPU版：
目前opencv中做h264编码实际上是向下调用ffmpeg的，而ffmpeg中则又是调用libx264，于是这也就涉及到了libx264，ffmpeg和opencv三者之间的联合编译
一、libx264
我用的系统是centos6.8，系统内默认安装的有libx264的，但是版本比较低，与ffmpeg不兼容，这里需要先卸载掉原先的x264再安装较新的版本，我用的版本是0.148
二、ffmpeg
系统同样默认安装了低版本的ffmpeg，这里也需要提前卸载掉，然后安装新版本，我用的版本是3.1.3
configuration: --prefix=/home/softwares/ffmpeg-3.1.3 --disable-yasm --enable-shared --enable-pic --enable-vaapi --enable-gpl --enable-libx264 --extra-cflags=-I/usr/local/x264/include --extra-ldflags=-L/usr/local/x264/lib
这个--enable-libx264很重要！！！
三、opencv
opencv我用的版本是2.4.13，别的版本（差的不多的话）应该都是好用的
configuration:
cmake -D CMAKE_BUILD_TYPE=release -D CMAKE_INSTALL_PREFIX=/home/softwares/opencv-2.4.13 -D WITH_CUDA=ON -D WITH_NVCUVID=ON ..
我们一般在调用VideoWriter的时候，都会用X264的FOURCC
writer.open("output_cpu.mp4", CV_FOURCC('X', '2', '6', '4'), FPS, frame.size());但是会有如下错误
[libx264 @ 0x8d6220] broken ffmpeg default settings detected
[libx264 @ 0x8d6220] use an encoding preset (e.g. -vpre medium)
[libx264 @ 0x8d6220] preset usage: -vpre -vpre
[libx264 @ 0x8d6220] speed presets are listed in x264 --help
[libx264 @ 0x8d6220] profile is optional; x264 defaults to high /* Detect default ffmpeg settings and terminate with an error. */
{
int score = 0;
score += h->param.analyse.i_me_range == 0;
score += h->param.rc.i_qp_step == 3;
score += h->param.i_keyint_max == 12;
score += h->param.rc.i_qp_min == 2;
score += h->param.rc.i_qp_max == 31;
score += h->param.rc.f_qcompress == 0.5;
score += fabs(h->param.rc.f_ip_factor - 1.25) < 0.01;
score += fabs(h->param.rc.f_pb_factor - 1.25) < 0.01;
score += h->param.analyse.inter == 0 && h->param.analyse.i_subpel_refine == 8;
if( score >= 5 )
{
x264_log( h, X264_LOG_ERROR, "broken ffmpeg default settings detected\n" );
x264_log( h, X264_LOG_ERROR, "use an encoding preset (vpre)\n" );
return -1;
}
}We must at least set 4 param of the AVCodecContext before open it.
/*default settings for x264*/
ctx->me_range = 16;
ctx->max_qdiff = 4;
ctx->qmin = 10;
ctx->qmax = 51;
ctx->qcompress = 0.6;
于是我们找到opencv根目录下，modules/highgui/src/cap_ffmpeg_impl.hpp找到相应调用ffmpeg的模块，在153~157行加上那些默认参数
bool CvVideoWriter_FFMPEG::open( const char * filename, int fourcc,
double fps, int width, int height, bool is_color )
{
CV_CODEC_ID codec_id = CV_CODEC(CODEC_ID_NONE);
int err, codec_pix_fmt;
double bitrate_scale = 1;
close();
// check arguments
if( !filename )
return false;
if(fps <= 0)
return false;
// we allow frames of odd width or height, but in this case we truncate
// the rightmost column/the bottom row. Probably, this should be handled more elegantly,
// but some internal functions inside FFMPEG swscale require even width/height.
width &= -2;
height &= -2;
if( width <= 0 || height <= 0 )
return false;
/* auto detect the output format from the name and fourcc code. */
#if LIBAVFORMAT_BUILD >= CALC_FFMPEG_VERSION(53, 2, 0)
fmt = av_guess_format(NULL, filename, NULL);
#else
fmt = guess_format(NULL, filename, NULL);
#endif
if (!fmt)
return false;
/* determine optimal pixel format */
if (is_color) {
input_pix_fmt = AV_PIX_FMT_BGR24;
}
else {
input_pix_fmt = AV_PIX_FMT_GRAY8;
}
/* Lookup codec_id for given fourcc */
#if LIBAVCODEC_VERSION_INT<((51<<16)+(49<<8)+0)
if( (codec_id = codec_get_bmp_id( fourcc )) == CV_CODEC(CODEC_ID_NONE) )
return false;
#else
const struct AVCodecTag * tags[] = { codec_bmp_tags, NULL};
if( (codec_id = av_codec_get_id(tags, fourcc)) == CV_CODEC(CODEC_ID_NONE) )
return false;
#endif
// alloc memory for context
#if LIBAVFORMAT_BUILD >= CALC_FFMPEG_VERSION(53, 2, 0)
oc = avformat_alloc_context();
#else
oc = av_alloc_format_context();
#endif
assert (oc);
/* set file name */
oc->oformat = fmt;
snprintf(oc->filename, sizeof(oc->filename), "%s", filename);
/* set some options */
oc->max_delay = (int)(0.7*AV_TIME_BASE); /* This reduces buffer underrun warnings with MPEG */
// set a few optimal pixel formats for lossless codecs of interest..
switch (codec_id) {
#if LIBAVCODEC_VERSION_INT>((50<<16)+(1<<8)+0)
case CV_CODEC(CODEC_ID_JPEGLS):
// BGR24 or GRAY8 depending on is_color...
codec_pix_fmt = input_pix_fmt;
break;
#endif
case CV_CODEC(CODEC_ID_HUFFYUV):
codec_pix_fmt = AV_PIX_FMT_YUV422P;
break;
case CV_CODEC(CODEC_ID_MJPEG):
case CV_CODEC(CODEC_ID_LJPEG):
codec_pix_fmt = AV_PIX_FMT_YUVJ420P;
bitrate_scale = 3;
break;
case CV_CODEC(CODEC_ID_RAWVIDEO):
codec_pix_fmt = input_pix_fmt == AV_PIX_FMT_GRAY8 ||
input_pix_fmt == AV_PIX_FMT_GRAY16LE ||
input_pix_fmt == AV_PIX_FMT_GRAY16BE ? input_pix_fmt : AV_PIX_FMT_YUV420P;
break;
default:
// good for lossy formats, MPEG, etc.
codec_pix_fmt = AV_PIX_FMT_YUV420P;
break;
}
double bitrate = MIN(bitrate_scale*fps*width*height, (double)INT_MAX/2);
// TODO -- safe to ignore output audio stream?
video_st = icv_add_video_stream_FFMPEG(oc, codec_id,
width, height, (int)(bitrate + 0.5),
fps, codec_pix_fmt);
/* set the output parameters (must be done even if no
parameters). */
#if LIBAVFORMAT_BUILD < CALC_FFMPEG_VERSION(53, 2, 0)
if (av_set_parameters(oc, NULL) < 0) {
return false;
}
#endif
#if 0
#if FF_API_DUMP_FORMAT
dump_format(oc, 0, filename, 1);
#else
av_dump_format(oc, 0, filename, 1);
#endif
#endif
/* now that all the parameters are set, we can open the audio and
video codecs and allocate the necessary encode buffers */
if (!video_st){
return false;
}
AVCodec *codec;
AVCodecContext *c;
#if LIBAVFORMAT_BUILD > 4628
c = (video_st->codec);
#else
c = &(video_st->codec);
#endif
c->codec_tag = fourcc;
/* find the video encoder */
codec = avcodec_find_encoder(c->codec_id);
if (!codec) {
fprintf(stderr, "Could not find encoder for codec id %d: %s", c->codec_id, icvFFMPEGErrStr(
#if LIBAVFORMAT_BUILD >= CALC_FFMPEG_VERSION(53, 2, 0)
AVERROR_ENCODER_NOT_FOUND
#else
-1
#endif
));
return false;
}
int64_t lbit_rate = (int64_t)c->bit_rate;
lbit_rate += (bitrate / 2);
lbit_rate = std::min(lbit_rate, (int64_t)INT_MAX);
c->bit_rate_tolerance = (int)lbit_rate;
c->bit_rate = (int)lbit_rate;
c->me_range = 16;
c->max_qdiff = 4;
c->qcompress = 0.6;
c->qmin = 10;
c->qmax = 51;
/* open the codec */
if ((err=
#if LIBAVCODEC_VERSION_INT >= ((53<<16)+(8<<8)+0)
avcodec_open2(c, codec, NULL)
#else
avcodec_open(c, codec)
#endif
) < 0) {
fprintf(stderr, "Could not open codec '%s': %s", codec->name, icvFFMPEGErrStr(err));
return false;
}
outbuf = NULL;
if (!(oc->oformat->flags & AVFMT_RAWPICTURE)) {
/* allocate output buffer */
/* assume we will never get codec output with more than 4 bytes per pixel... */
outbuf_size = width*height*4;
outbuf = (uint8_t *) av_malloc(outbuf_size);
}
bool need_color_convert;
need_color_convert = (c->pix_fmt != input_pix_fmt);
/* allocate the encoded raw picture */
picture = icv_alloc_picture_FFMPEG(c->pix_fmt, c->width, c->height, need_color_convert);
if (!picture) {
return false;
}
/* if the output format is not our input format, then a temporary
picture of the input format is needed too. It is then converted
to the required output format */
input_picture = NULL;
if ( need_color_convert ) {
input_picture = icv_alloc_picture_FFMPEG(input_pix_fmt, c->width, c->height, false);
if (!input_picture) {
return false;
}
}
/* open the output file, if needed */
if (!(fmt->flags & AVFMT_NOFILE)) {
#if LIBAVFORMAT_BUILD < CALC_FFMPEG_VERSION(53, 2, 0)
if (url_fopen(&oc->pb, filename, URL_WRONLY) < 0)
#else
if (avio_open(&oc->pb, filename, AVIO_FLAG_WRITE) < 0)
#endif
{
return false;
}
}
#if LIBAVFORMAT_BUILD >= CALC_FFMPEG_VERSION(52, 111, 0)
/* write the stream header, if any */
err=avformat_write_header(oc, NULL);
#else
err=av_write_header( oc );
#endif
if(err < 0)
{
close();
remove(filename);
return false;
}
frame_width = width;
frame_height = height;
frame_idx = 0;
ok = true;
return true;
}
CvCapture_FFMPEG* cvCreateFileCapture_FFMPEG( const char* filename )
{
CvCapture_FFMPEG* capture = (CvCapture_FFMPEG*)malloc(sizeof(*capture));
capture->init();
if( capture->open( filename ))
return capture;
capture->close();
free(capture);
return 0;
}
这里给AVCodecContext加入相应参数即可，接着重新编译OpenCV，在调用VideoWriter即可成功啦
GPU版：
在OpenCV较高的版本中，OpenCV提供了CUDA的h264编码相关的API，然而不幸的是
取而代之的是NVIDIA提供了自己的VideoCodecSDK，在CUDA Developer的网站里可以下载到相关的SDK的，当然这里对GPU的版本也有一些限制啦
使用的GPU架构至少是GM10X以上的
当然对CUDA的驱动也是有要求的，具体的文档里也是有说明的