我做了一些关于减少视频延迟的研究。
我的following answer演示了相关的FFmpeg标志是-probesize 32和-flags low_delay。
上述标志与视频解码器侧（接收器侧）有关。
视频编码参数"发送器/编码器侧"对于确定端到端延迟更重要。
添加参数-tune zerolatency可以将编码器延迟降至最低，但所需的带宽要高得多（可能与Internet上的流媒体无关）。
我将把我的答案限制在解码延迟上，因为它似乎与您的问题主题更相关。
关于"知道其他人如何获得低延迟的视频帧"的主题是一个单独问题的主题（我不知道答案）。
为了比较FFplay和FFmpeg（解码器）之间的延迟差异，我创建了一个"自包含"测试示例。
主要"原则"：

• 执行FFmpeg子进程，以并行流式传输两个RTSP输出流。

流视频是合成模式，帧计数器作为视频上的文本。
两个输出流应用相同的编码参数（仅端口不同）。
RTSP IP地址为127.0.0.1（本地主机）。
(Note：我们可能会使用tee muxer而不是编码两次，但我从来没有尝试过）。

• 执行FFplay子进程，解码并显示一个视频流。
• 执行FFmpeg子进程，以解码其他视频流。

OpenCV imshow用于显示视频。

• 具有较大计数器的显示视频是具有较低延迟的视频。

代码示例（已更新）：

import cv2
import numpy as np
import subprocess as sp
import shlex

rtsp_stream0 = 'rtsp://127.0.0.1:21415/live.stream'  # Use localhost for testing 
rtsp_stream1 = 'rtsp://127.0.0.1:31415/live.stream'
width = 256  # Use low resolution (for testing).
height = 144
fps = 30

# https://stackoverflow.com/questions/60462840/ffmpeg-delay-in-decoding-h264
ffmpeg_cmd = shlex.split(f'ffmpeg -nostdin -probesize 32 -flags low_delay -fflags nobuffer -rtsp_flags listen -rtsp_transport tcp -stimeout 1000000 -an -i {rtsp_stream0} -pix_fmt bgr24 -an -vcodec rawvideo -f rawvideo pipe:')

# FFplay command before updating the code (latency is still too high):  
# ffplay_cmd = shlex.split(f'ffplay -probesize 32 -analyzeduration 0 -sync ext -fflags nobuffer -flags low_delay -avioflags direct -rtsp_flags listen -strict experimental -framedrop -rtsp_transport tcp -listen_timeout 1000000 {rtsp_stream1}')

# Updated FFplay command - adding "-vf setpts=0" (fixing the latency issue):
# https://stackoverflow.com/questions/16658873/how-to-minimize-the-delay-in-a-live-streaming-with-ffmpeg
ffplay_cmd = shlex.split(f'ffplay -probesize 32 -analyzeduration 0 -sync ext -fflags nobuffer -flags low_delay -avioflags direct -rtsp_flags listen -strict experimental -framedrop -vf setpts=0 -rtsp_transport tcp -listen_timeout 1000000 {rtsp_stream1}')

# Execute FFplay to used as reference
ffplay_process = sp.Popen(ffplay_cmd)

# Open sub-process that gets in_stream as input and uses stdout as an output PIPE.
process = sp.Popen(ffmpeg_cmd, stdout=sp.PIPE) #,stderr=sp.DEVNULL

# The following FFmpeg sub-process stream RTSP video.
# The video is synthetic video with frame counter (that counts every frame) at 30fps.
# The arguments of the encoder are almost default arguments - not tuned for low latency.
# drawtext filter with the n or frame_num function https://stackoverflow.com/questions/15364861/frame-number-overlay-with-ffmpeg
rtsp_streaming_process = sp.Popen(shlex.split(f'ffmpeg -re -f lavfi -i testsrc=size={width}x{height}:rate={fps} '
                                               '-filter_complex "drawtext=fontfile=Arial.ttf: text=''%{frame_num}'': start_number=1: x=(w-tw)/2: y=h-(2*lh): fontcolor=black: fontsize=72: box=1: boxcolor=white: boxborderw=5",'
                                               'split[v0][v1] '  # Split the input into [v0] and [v1]
                                               '-vcodec libx264 -pix_fmt yuv420p -g 30 -rtsp_transport tcp -f rtsp -muxdelay 0.1 -bsf:v dump_extra '
                                              f'-map "[v0]" -an {rtsp_stream0} '
                                               '-vcodec libx264 -pix_fmt yuv420p -g 30 -rtsp_transport tcp -f rtsp -muxdelay 0.1 -bsf:v dump_extra '
                                              f'-map "[v1]" -an {rtsp_stream1}'))

while True:
    raw_frame = process.stdout.read(width*height*3)

    if len(raw_frame) != (width*height*3):
        print('Error reading frame!!!')  # Break the loop in case of an error (too few bytes were read).
        break

    # Transform the byte read into a numpy array, and reshape it to video frame dimensions
    frame = np.frombuffer(raw_frame, np.uint8)
    frame = frame.reshape((height, width, 3))

    # Show frame for testing
    cv2.imshow('frame', frame)
    key = cv2.waitKey(1)

    if key == 27:
        break
  
process.stdout.close()
process.wait()
ffplay_process.kill()
rtsp_streaming_process.kill()
cv2.destroyAllWindows()

添加-vf setpts=0之前的示例输出：
样本输出（左侧为OpenCV，右侧为FFplay）：

在将-vf setpts=0添加到FFplay命令之前，FFmpeg-OpenCV延迟似乎降低了6帧。
注意：我花了一些时间才找到解决方案，我决定保留原帖子的结果，以显示添加setpts过滤器的重要性。

更新：

添加-vf setpts=0解决了延迟问题。
following post的最新回答建议添加setpts视频过滤器，将所有视频时间戳重置为零。
这可能不是一个好主意与目前的音频流，但当最低的视频延迟是必需的，这是最好的解决方案，我可以找到。
添加-vf setpts=0后，FFplay和OpenCV的延迟大致相同：

使用mpv media player重复测试：
(Note：在我找到FFplay解决方案之前似乎更相关）。
当应用this page中的所有mpv "延迟破解"时，mpv和OpenCV的延迟大致相同：

FFplay一定有解决的办法，但我找不到......
代码示例（使用mpv而不是FFplay）：

import cv2
import numpy as np
import subprocess as sp
import shlex

rtsp_stream0 = 'rtsp://127.0.0.1:21415/live.stream'  # Use localhost for testing 
rtsp_stream1 = 'rtsp://127.0.0.1:31415/live.stream'
width = 256  # Use low resolution (for testing).
height = 144
fps = 30

# https://stackoverflow.com/questions/60462840/ffmpeg-delay-in-decoding-h264
ffmpeg_cmd = shlex.split(f'ffmpeg -nostdin -probesize 32 -flags low_delay -fflags nobuffer -rtsp_flags listen -rtsp_transport tcp -stimeout 1000000 -an -i {rtsp_stream0} -pix_fmt bgr24 -an -vcodec rawvideo -f rawvideo pipe:')

# https://stackoverflow.com/questions/16658873/how-to-minimize-the-delay-in-a-live-streaming-with-ffmpeg
#ffplay_cmd = shlex.split(f'ffplay -probesize 32 -analyzeduration 0 -sync ext -fflags nobuffer -flags low_delay -avioflags direct -rtsp_flags listen -strict experimental -framedrop -rtsp_transport tcp -listen_timeout 1000000 {rtsp_stream1}')

# https://github.com/mpv-player/mpv/issues/4213
mpv_cmd = shlex.split(f'mpv --demuxer-lavf-o=rtsp_flags=listen --rtsp-transport=tcp --profile=low-latency --no-cache --untimed --no-demuxer-thread --vd-lavc-threads=1 {rtsp_stream1}')

# Execute FFplay to used as reference
#ffplay_process = sp.Popen(ffplay_cmd)

# Execute mpv media player (as reference)
mpv_process = sp.Popen(mpv_cmd)

# Open sub-process that gets in_stream as input and uses stdout as an output PIPE.
process = sp.Popen(ffmpeg_cmd, stdout=sp.PIPE) #,stderr=sp.DEVNULL

# The following FFmpeg sub-process stream RTSP video.
# The video is synthetic video with frame counter (that counts every frame) at 30fps.
# The arguments of the encoder are almost default arguments - not tuned for low latency.
# drawtext filter with the n or frame_num function https://stackoverflow.com/questions/15364861/frame-number-overlay-with-ffmpeg
rtsp_streaming_process = sp.Popen(shlex.split(f'ffmpeg -re -f lavfi -i testsrc=size={width}x{height}:rate={fps} '
                                               '-filter_complex "drawtext=fontfile=Arial.ttf: text=''%{frame_num}'': start_number=1: x=(w-tw)/2: y=h-(2*lh): fontcolor=black: fontsize=72: box=1: boxcolor=white: boxborderw=5",'
                                               'split[v0][v1] '  # Split the input into [v0] and [v1]
                                               '-vcodec libx264 -pix_fmt yuv420p -g 30 -rtsp_transport tcp -f rtsp -muxdelay 0.1 -bsf:v dump_extra '
                                              f'-map "[v0]" -an {rtsp_stream0} '
                                               '-vcodec libx264 -pix_fmt yuv420p -g 30 -rtsp_transport tcp -f rtsp -muxdelay 0.1 -bsf:v dump_extra '
                                              f'-map "[v1]" -an {rtsp_stream1}'))

while True:
    raw_frame = process.stdout.read(width*height*3)

    if len(raw_frame) != (width*height*3):
        print('Error reading frame!!!')  # Break the loop in case of an error (too few bytes were read).
        break

    # Transform the byte read into a numpy array, and reshape it to video frame dimensions
    frame = np.frombuffer(raw_frame, np.uint8)
    frame = frame.reshape((height, width, 3))

    # Show frame for testing
    cv2.imshow('frame', frame)
    key = cv2.waitKey(1)

    if key == 27:
        break
  
process.stdout.close()
process.wait()
#ffplay_process.kill()
mpv_process.kill()
rtsp_streaming_process.kill()
cv2.destroyAllWindows()

假设瓶颈确实在示例代码中的某个地方（而不是在<do stuff with frame/ show frame etc.>中），您可以尝试更新numpy array，而不是每次都创建一个：

frame = np.empty((HEIGHT, WIDTH, 3), np.uint8) 
frame_bytes = memoryview(frame).cast("b")
while True:
    process.stdout.readinto(frame_bytes) # fills the buffer of frame
    ...