System information (version)

• OpenCV => 4.5.2
• Operating System / Platform => Linux x64
• Compiler => GCC
• API: Python 3.8.9
• nVidia Cuda: 11.2.1
• Cudnn: 8.1.1

Detailed description

After upgrading from 4.5.1 to 4.5.2 the ffmpeg backend videoio decoding/encoding broke.
Also posted in the forum:
https://forum.opencv.org/t/opencv-4-5-2-breaking-change-with-ffmpeg-hw-accelerated-decoding/2613/2

Steps to reproduce

Old syntax to start a video capture stream

os.environ["OPENCV_FFMPEG_CAPTURE_OPTIONS"]="hwaccel;cuvid|video_codec;h264_cuvid|vsync;0"
stream = cv2.VideoCapture(rtsp_server_uri, cv2.CAP_FFMPEG)

Upgrading to 4.5.2 will lead to failure to start the stream without any error message on python.

After looking at documentation it appears that there are new features additions which have broken the hardware acceleration. After a few tests, the closest I have come to get it to work was:

os.environ[ "OPENCV_FFMPEG_CAPTURE_OPTIONS"] = "hw_decoders_any;cuda"
stream = cv2.VideoCapture(rtsp_server_uri, cv2.CAP_FFMPEG, (cv2.CAP_PROP_HW_ACCELERATION, cv2.VIDEO_ACCELERATION_ANY))

But it still found the url and started the stream without any decoding. Verified the problem by recompiling 4.5.1 with the exact same libraries and found it to work fine.