在計算機視覺領域中,視頻算法是重要的一個部分,不一樣於圖像,視頻須要含有時序特徵的多幀圖像,同時,還包括必定的運動信息,如光流。在預處理時,須要將視頻中的圖像和光流提取出來,開源工程dense_flow已經實現這個功能,支持GPU操做。python
在CUDA 9和OpenCV 3中,配置dense_flow工程,高級版本temporal-segment-networks。同時,推薦視頻的Benchmark工程mmaction。git
參考:github
dense_flow: https://github.com/yjxiong/dense_flow
temporal-segment-networks: https://github.com/yjxiong/temporal-segment-networks
mmaction: https://github.com/open-mmlab/mmaction
複製代碼
OpenCV的編譯步驟以下:算法
CUDA 9不支持OpenCV2.x,只能選用3.x,如3.1.0,參考ubuntu
下載OpenCV源碼文件,並解壓:bash
wget https://github.com/opencv/opencv/archive/3.1.0.zip
unzip 3.1.0.zip
cd opencv-3.1.0
複製代碼
在opencv-3.1.0中,下載opencv_contrib文件,並解壓:ide
wget https://github.com/opencv/opencv_contrib/archive/3.1.0.zip
unzip 3.1.0.zip
複製代碼
位置以下:測試
緣由是,SURF或SIFT算法移入opencv_contrib
,須要參於源碼編譯,在dense_flow
中,調用SURF算法,不然沒法找到SURF,參考。ui
Error:this
undefined reference to `cv::xfeatures2d::SURF::create(double, int, int, bool, bool)'
複製代碼
CMake Error:
CMake Error: The following variables are used in this project, but they are set to NOTFOUND.
Please set them or make sure they are set and tested correctly in the CMake files:
CUDA_nppi_LIBRARY (ADVANCED)
複製代碼
緣由是,nppi已經廢棄,須要替換其餘的CUDA,同時,CUDA 2.0已經不兼容當前版本,須要刪除。
須要修改cmake文件夾下的FindCUDA.cmake和OpenCVDetectCUDA.cmake,還有修改common.hpp。
修改FindCUDA.cmake文件,3處替換:
替換
find_cuda_helper_libs(nppi)
複製代碼
爲
find_cuda_helper_libs(nppial)
find_cuda_helper_libs(nppicc)
find_cuda_helper_libs(nppicom)
find_cuda_helper_libs(nppidei)
find_cuda_helper_libs(nppif)
find_cuda_helper_libs(nppig)
find_cuda_helper_libs(nppim)
find_cuda_helper_libs(nppist)
find_cuda_helper_libs(nppisu)
find_cuda_helper_libs(nppitc)
複製代碼
替換
set(CUDA_npp_LIBRARY "${CUDA_nppc_LIBRARY};${CUDA_nppi_LIBRARY};${CUDA_npps_LIBRARY}")
複製代碼
爲
set(CUDA_npp_LIBRARY "${CUDA_nppc_LIBRARY};${CUDA_nppial_LIBRARY};${CUDA_nppicc_LIBRARY};${CUDA_nppicom_LIBRARY};${CUDA_nppidei_LIBRARY};${CUDA_nppif_LIBRARY};${CUDA_nppig_LIBRARY};${CUDA_nppim_LIBRARY};${CUDA_nppist_LIBRARY};${CUDA_nppisu_LIBRARY};${CUDA_nppitc_LIBRARY};${CUDA_npps_LIBRARY}")
複製代碼
替換
unset(CUDA_nppi_LIBRARY CACHE)
複製代碼
爲
unset(CUDA_nppial_LIBRARY CACHE)
unset(CUDA_nppicc_LIBRARY CACHE)
unset(CUDA_nppicom_LIBRARY CACHE)
unset(CUDA_nppidei_LIBRARY CACHE)
unset(CUDA_nppif_LIBRARY CACHE)
unset(CUDA_nppig_LIBRARY CACHE)
unset(CUDA_nppim_LIBRARY CACHE)
unset(CUDA_nppist_LIBRARY CACHE)
unset(CUDA_nppisu_LIBRARY CACHE)
unset(CUDA_nppitc_LIBRARY CACHE)
複製代碼
修改OpenCVDetectCUDA.cmake文件,2處刪除:
將"Fermi"註釋,將"Kepler"提早,即刪除"Fermi"的if分支,主要是爲了刪除CUDA的2.0版本兼容。
set(__cuda_arch_ptx "")
if(CUDA_GENERATION STREQUAL "Fermi")
set(__cuda_arch_bin "2.0")
elseif(CUDA_GENERATION STREQUAL "Kepler")
set(__cuda_arch_bin "3.0 3.5 3.7")
複製代碼
修改成
set(__cuda_arch_ptx "")
if(CUDA_GENERATION STREQUAL "Kepler")
set(__cuda_arch_bin "3.0 3.5 3.7")
複製代碼
在CUDA版本大於6.5時,刪除2.0版本的兼容,修改完成以下:
elseif(${CUDA_VERSION} VERSION_GREATER "6.5")
set(__cuda_arch_bin "3.0 3.5")
複製代碼
在opencv-3.1.0/modules/cudev/include/opencv2/cudev/common.hpp
的頭文件中,添加:
#include <cuda_fp16.h>
複製代碼
Error:
hdf5.hpp:40:18: fatal error: hdf5.h: No such file or directory
複製代碼
修改opencv-3.1.0/modules/python/common.cmake文件,在文件頭部中,添加
find_package(HDF5)
include_directories(${HDF5_INCLUDE_DIRS})
複製代碼
Error:
fatal error: opencv2/nonfree/nonfree.hpp: No such file or directory
複製代碼
安裝包libopencv-nonfree-dev:
sudo apt-get update
sudo add-apt-repository --yes ppa:xqms/opencv-nonfree
sudo apt-get update
sudo apt-get install libopencv-nonfree-dev
複製代碼
若是不成功,更換ppa的源:
sudo add-apt-repository --remove ppa:xqms/opencv-nonfree
sudo add-apt-repository --yes ppa:jeff250/opencv
sudo apt-get update
sudo apt-get install libopencv-dev
sudo apt-get install libopencv-nonfree-dev
複製代碼
Error:
c->flags |= CODEC_FLAG_GLOBAL_HEADER
複製代碼
在opencv-3.1.0/modules/videoio/src/cap_ffmpeg_impl.hpp
中,添加:
#define AV_CODEC_FLAG_GLOBAL_HEADER (1 << 22)
#define CODEC_FLAG_GLOBAL_HEADER AV_CODEC_FLAG_GLOBAL_HEADER
#define AVFMT_RAWPICTURE 0x0020
複製代碼
執行make操做,在OPENCV_EXTRA_MODULES_PATH
中,須要引入opencv_contrib
:
make -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local -D OPENCV_EXTRA_MODULES_PATH=/data1/wcl/workspace/opencv-3.1.0/opencv_contrib-3.1.0/modules/ ..
複製代碼
執行make,32進程:
make -j32
複製代碼
安裝,而且將opencv導入系統環境。
sudo make install
sudo /bin/bash -c 'echo "/usr/local/lib" > /etc/ld.so.conf.d/opencv.conf'
sudo ldconfig
複製代碼
安裝libzip-dev
apt-get install libzip-dev
複製代碼
下載dense_flow工程,切換OpenCV的3.1分支:
git clone --recursive http://github.com/yjxiong/dense_flow
git checkout remotes/origin/opencv-3.1
複製代碼
指定OpenCV_DIR,編譯工程:
mkdir build && cd build
OpenCV_DIR=/opencv-3.1.0/build cmake .. -DCUDA_USE_STATIC_CUDA_RUNTIME=OFF
make -j32
複製代碼
編譯成功以後,在build文件夾中:
flow_x
前綴,y軸光流以flow_y
前綴,其他參數參考。./extract_gpu -f=980044841.mp4 -x=./tmp/flow_x -y=./tmp/flow_y -i=./tmp/image -b=20 -t=1 -d=0 -s=1 -o=dir
複製代碼
Error,提示沒法打開視頻,將空格替換爲「=」便可。
FATAL [default] Check failed: [video_stream.isOpened()]
複製代碼
測試視頻:
輸出結果:
GPU使用狀況
OK, that's all! Enjoy it!