This is merely a quick summary of the points that differ from another older guide, that handles the same topic:
There is no need to change the system path or set PKG_CONFIG_PATH as indicated in the linked article. Instead set the environment variable GSTREAMER_DIR to the right directory (in my case it is: D:\gstreamer\1.0\x86_64 ), and append D:\gstreamer\1.0\x86_64\bin to the PATH environment variable (not the \lib or \lib\gstreamer-1.0 folders).
cap_gstreamer.cpp does not need to be modified starting with OpenCV 4. It is fully compatible with MSVC 2017 and Windows. (Earlier versions of OpenCV and MSVC may work as well, but I haven’t tested them.)
It is not necessary to change opencv\sources\CMakeLists.txt anymore. Therefore there is no need to download, install or add pkg-config to the system path, either.
OpenCV contains now a file FindGstreamerWindows.cmake that will locate GStreamer, but you still have to set the environment variable GSTREAMER_DIR properly, as mentioned above.
Make sure to select the x64 compiler toolset (or the x86 one), depending on which you need. CMake will only generate projects for one of them.
For example if you use 64 bit Python, then OpenCV should be 64 bit as well.
Search for “gstreamer” and “python” in the CMake GUI, and check that all variables could be filled out correctly, and that GStreamer is enabled. Also search for examples and check BUILD_EXAMPLES, which allows for easier debugging in case something went wrong (see “Resolving issues” below). You might also want to check INSTALL_PYTHON_EXAMPLES, for the same reason.
After you configured everything and no red entries are left, then generated the project files, load the solution in VS2017. Select the Release config, build the ALL_BUILD project, then activate the INSTALL project, and build it to install the Python bindings.
Ideally all these changes would be part of OpenCV itself, currently it looks like a hack specific to GStreamer, which is why I doubt they would accept it as a patch, unchanged.
Sometimes there can be errors when loading OpenCV and GStreamer from Python, due to path issues, or having provided the wrong libraries to OpenCV, etc.
As a general advice: make sure there are no wrong / old paths set in the environment variable PATH, or other environment variables, such as GSTREAMER_DIR, that point to directories of previous OpenCV / GStreamer files.
(The debugging procedure below is pretty specific to my needs, it should be possible to find a more generic, and easier to follow procedure for everyone, if there is any interest.)
As a more targeted way to trace down the cause, a first good step is to see if GStreamer works from OpenCV in C++. If you successfully compiled OpenCV, open the project samples/cpp/videocapture_basic. From there, open the file videocapture_basic.cpp and find this line:
cap.open(deviceID + apiID);
Comment it out, and append these two lines:
cv::String gst_str = "udpsrc port=5200 ! gdpdepay ! rtph264depay ! avdec_h264 ! videoconvert ! appsink sync=false"; cap.open(gst_str, cv::CAP_GSTREAMER);
Select videocapture_basic as the start project, then run. It should patiently wait for a stream coming from GStreamer. From another computer, in my case a Raspberry Pi in the local network, provide the stream by executing this console command:
gst-launch-1.0 -v rpicamsrc preview=0 bitrate=2000000 sensor_mode=6 vflip=true ! h264parse ! rtph264pay config-interval=1 pt=96 ! gdppay ! udpsink host=192.168.1.50 port=5200
If a window pops up that displays the video, OpenCV was compiled properly, and GStreamer works well. The next step is doing the same test with the equivalent Python sample.
A likely cause for errors is to have the PATH set up wrong or having started the wrong Python environment / Anaconda / Spyder version (possibly because start menu links are wrong or outdated from other installs). It might also be that Python / Anaconda modify the PATH variable or use other strategies to load DLLs which ultimately cause it to load the wrong ones.
Resolving issues such as loading DLLs that are missing imports, or DLLs that cannot be found, is a lot easier to debug from C++. So, for this kind of error, it should be tried to figure it out with the simple sample project videocapture_basic, with the slight modifications as mentioned above.