python按照指定时间范围裁剪视频

按照指定时间裁剪视频通常有如下三种方法：

• 使用 moviepy 库的 clip.subclip(start_time, stop_time) 方法实现：Clips transformations and effects — MoviePy 1.0.2 documentation (zulko.github.io)
• 使用 opencv-python 库的 cap.set(cv2.CAP_PROP_POS_MSEC, float(start_time * MILLISECONDS))方法实现：Reading and Writing Images and Video — OpenCV 2.4.13.7 documentation
• 使用 ffmpeg-python 库的 trim 方法实现：ffmpeg-python: Python bindings for FFmpeg — ffmpeg-python documentation (kkroening.github.io)

但对于从监控等视频流保存的视频文件来说，可能存在文件的起始时间戳不准确的情况。即第一帧视频数据对应的时间戳不是 00:00:00 。这种情况下，上述方法的前两种将无法正确处理（第三种方法未尝试）。因此，我采用了一种更笨（速度更慢，但更直观）的做法。即逐帧读取视频，通过判断各帧与第一帧的相对时间来确定是否保留当前帧，由此规避了首帧时间戳错误的问题。程序片段记录如下：

import cv2
 
def clip_video_with_opencv(src_path:str, dst_path:str, start_time:str, stop_time:str):
    """ Clip the video with OpenCV for the specified time period (Force the start timestamp of the video to `00:00:00`)
 
    This function is time consuming.
    If the start timestamp of the video is correct, you can use the `cap.set()` function to jump to the 
    start time of the target clip directly, or use the `moviepy` module instead.
 
    Args:
        `src_path` (str): Source video path. The path of the video to be clipped
        `dst_path` (str): Destination video path. Storage path of the clipped video
        `start_time` (str): Time to start clipping. The format is `hh:mm:ss`
        `stop_time` (str): Time to stop clipping. `None` if the stop time is the end of the video. The format is `hh:mm:ss`
    """
 
    th, tm, ts = map(int, start_time.split(":"))
    start_time = th*3600 + tm*60 + ts  # Seconds
    if stop_time:
        th, tm, ts = map(int, stop_time.split(":"))
        stop_time = th*3600 + tm*60 + ts  # Seconds
 
    ### Prepare to load video data
    in_cap = cv2.VideoCapture(src_path)
    frame_total = int(in_cap.get(cv2.CAP_PROP_FRAME_COUNT))
    fps = int(in_cap.get(cv2.CAP_PROP_FPS))
    # fourcc_int = int(in_cap.get(cv2.CAP_PROP_FOURCC))
    frame_width, frame_height = int(in_cap.get(cv2.CAP_PROP_FRAME_WIDTH)), int(in_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
 
    ### Prepare to save video clip
    # fourcc_str = "{}{}{}{}".format(chr(fourcc_int&255), chr((fourcc_int>>8)&255), chr((fourcc_int>>16)&255), chr((fourcc_int>>24)&255))
    # fourcc = cv2.VideoWriter_fourcc(*fourcc_str)
    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    out_cap = cv2.VideoWriter(dst_path, fourcc, fps, (frame_width, frame_height))
 
    ### Frame-by-frame processing
    frame_idx = 0
    start_ts = None
    while in_cap.isOpened():
        success = in_cap.grab()  # !!! Load a frame of data but do not decode it (decoding is time consuming)
        if not success:
            break
        video_ts = int(round(int(in_cap.get(cv2.CAP_PROP_POS_MSEC))/1000))  # Seconds
        if start_ts is None:
            start_ts = video_ts
        video_ts -= start_ts  # Force the start timestamp of the video to 00:00:00
        th, tm, ts = video_ts//3600, (video_ts%3600)//60, video_ts%3600%60
        if video_ts < start_time:  # Skip the video frames before the target clip
            print("{}/{}: {:0>2d}:{:0>2d}:{:0>2d}  [skip]".format(frame_idx+1, frame_total, th, tm, ts), end="\r")
        elif video_ts >= start_time and (stop_time is None or video_ts <= stop_time):  # Save video frames of the target clip
            print("{}/{}: {:0>2d}:{:0>2d}:{:0>2d} [write]".format(frame_idx+1, frame_total, th, tm, ts), end="\r")
            success, frame = in_cap.retrieve()  # !!! Decode the frame data. Only the video frames of the target clip are decoded
            if not success:
                break
            out_cap.write(frame)
        else:
            break
        frame_idx += 1
    in_cap.release()
    out_cap.release()
    print("{}/{}: {:0>2d}:{:0>2d}:{:0>2d}  [done]".format(frame_idx, frame_total, th, tm, ts))