据我所知，OpenCV不允许选择视频流，因此这是不可能的。但是，您可以使用ffmpeg命令行实用工具轻松地做到这一点：

import numpy as np
import json
import subprocess

def videoInfo(filename):
    proc = subprocess.run([
        *"ffprobe -v quiet -print_format json -show_format -show_streams".split(),filename
    ],capture_output=True)
    proc.check_returncode()
    return json.loads(proc.stdout)

def readVideo(filename):
    cmd = ["ffmpeg","-i",filename]
    streams = 0
    for stream in videoInfo(filename)["streams"]:
        index = stream["index"]
        if stream["codec_type"] == "video":
            width = stream["width"]
            height = stream["height"]
            cmd += "-map",f"0:{index}"
            streams = streams + 1
    cmd += "-f","rawvideo","-pix_fmt","rgb24","-"
    shape = np.array([streams,height,width,3])
    with subprocess.Popen(cmd,stdout=subprocess.PIPE) as proc:
        while True:
            data = proc.stdout.read(shape.prod())  # One byte per each element
            if not data:
                return
            yield np.frombuffer(data,dtype=np.uint8).reshape(shape)

请注意，该代码读取所有视频流，并假定每个视频流具有相同的分辨率。它缺乏适当的错误处理，但是在我的科学项目中完成了工作。

例如，读取立体流：

import matplotlib.pyplot as plt

for left,right in readVideo("testvideo.mkv"):
    plt.imshow(left)
    plt.show()
    plt.imshow(right)
    plt.show()