需要卸載之前的opencv-python版本
pip uninstall opencv-python
pip uninstall opencv-contrib-python
 
安裝新的版本
pip install opencv_python==3.4.2.16 
pip install opencv-contrib-python==3.4.2.16

import cv2
import numpy as np
from tqdm import tqdm
import argparse
import os
 
# get param
parser = argparse.ArgumentParser(description='')
parser.add_argument('-v', type=str, default='')  # 指定輸入視頻路徑位置（參數(shù)必選）
parser.add_argument('-o', type=str, default='')  # 指定輸出視頻路徑位置（參數(shù)必選）
parser.add_argument('-n', type=int, default=-1)  # 指定處理的幀數(shù)（參數(shù)可選）, 不設(shè)置使用視頻實(shí)際幀
 
# eg: python3 stable.py -v=video/01.mp4 -o=video/01_stable.mp4 -n=100 -p=6
 
args = parser.parse_args()
 
input_path = args.v
output_path = args.o
number = args.n
 
class Stable:
    # 處理視頻文件路徑
    __input_path = None
 
    __output_path = None
 
    __number = number
 
    # surf 特征提取
    __surf = {
        # surf算法
        'surf': None,
        # 提取的特征點(diǎn)
        'kp': None,
        # 描述符
        'des': None,
        # 過濾后的特征模板
        'template_kp': None
    }
 
    # capture
    __capture = {
        # 捕捉器
        'cap': None,
        # 視頻大小
        'size': None,
        # 視頻總幀
        'frame_count': None,
        # 視頻幀率
        'fps': None,
        # 視頻
        'video': None,
    }
 
    # 配置
    __config = {
        # 要保留的最佳特征的數(shù)量
        'key_point_count': 5000,
        # Flann特征匹配
        'index_params': dict(algorithm=0, trees=5),
        'search_params': dict(checks=50),
        'ratio': 0.5,
    }
 
    # 特征提取列表
    __surf_list = []
 
    def __init__(self):
        pass
 
    # 初始化capture
    def __init_capture(self):
        self.__capture['cap'] = cv2.VideoCapture(self.__video_path)
        self.__capture['size'] = (int(self.__capture['cap'].get(cv2.CAP_PROP_FRAME_WIDTH)),
                                  int(self.__capture['cap'].get(cv2.CAP_PROP_FRAME_HEIGHT)))
 
        self.__capture['fps'] = self.__capture['cap'].get(cv2.CAP_PROP_FPS)
 
        self.__capture['video'] = cv2.VideoWriter(self.__output_path, cv2.VideoWriter_fourcc(*"mp4v"),
                                                  self.__capture['fps'], self.__capture['size'])
 
        self.__capture['frame_count'] = int(self.__capture['cap'].get(cv2.CAP_PROP_FRAME_COUNT))
 
        if number == -1:
            self.__number = self.__capture['frame_count']
        else:
            self.__number = min(self.__number, self.__capture['frame_count'])
 
    # 初始化surf
    def __init_surf(self):
 
        self.__capture['cap'].set(cv2.CAP_PROP_POS_FRAMES, 0)
        state, first_frame = self.__capture['cap'].read()
 
        self.__capture['cap'].set(cv2.CAP_PROP_POS_FRAMES, self.__capture['frame_count'] - 1)
        state, last_frame = self.__capture['cap'].read()
 
        self.__surf['surf'] = cv2.xfeatures2d.SURF_create(self.__config['key_point_count'])
 
        self.__surf['kp'], self.__surf['des'] = self.__surf['surf'].detectAndCompute(first_frame, None)
        kp, des = self.__surf['surf'].detectAndCompute(last_frame, None)
 
        # 快速臨近匹配
        flann = cv2.FlannBasedMatcher(self.__config['index_params'], self.__config['search_params'])
        matches = flann.knnMatch(self.__surf['des'], des, k=2)
 
        good_match = []
        for m, n in matches:
            if m.distance < self.__config['ratio'] * n.distance:
                good_match.append(m)
 
        self.__surf['template_kp'] = []
        for f in good_match:
            self.__surf['template_kp'].append(self.__surf['kp'][f.queryIdx])
 
    # 釋放
    def __release(self):
        self.__capture['video'].release()
        self.__capture['cap'].release()
 
    # 處理
    def __process(self):
 
        current_frame = 1
 
        self.__capture['cap'].set(cv2.CAP_PROP_POS_FRAMES, 0)
 
        process_bar = tqdm(self.__number, position=current_frame)
 
        while current_frame <= self.__number:
            # 抽幀
            success, frame = self.__capture['cap'].read()
 
            if not success: return
 
            # 計(jì)算
            frame = self.detect_compute(frame)
 
            # 寫幀
            self.__capture['video'].write(frame)
 
            current_frame += 1
 
            process_bar.update(1)
 
    # 視頻穩(wěn)像
    def stable(self, input_path, output_path, number):
        self.__video_path = input_path
        self.__output_path = output_path
        self.__number = number
 
        self.__init_capture()
        self.__init_surf()
        self.__process()
        self.__release()
 
    # 特征點(diǎn)提取
    def detect_compute(self, frame):
 
        frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
 
        # 計(jì)算特征點(diǎn)
        kp, des = self.__surf['surf'].detectAndCompute(frame_gray, None)
 
        # 快速臨近匹配
        flann = cv2.FlannBasedMatcher(self.__config['index_params'], self.__config['search_params'])
        matches = flann.knnMatch(self.__surf['des'], des, k=2)
 
        # 計(jì)算單應(yīng)性矩陣
        good_match = []
        for m, n in matches:
            if m.distance < self.__config['ratio'] * n.distance:
                good_match.append(m)
 
        # 特征模版過濾
        p1, p2 = [], []
        for f in good_match:
            if self.__surf['kp'][f.queryIdx] in self.__surf['template_kp']:
                p1.append(self.__surf['kp'][f.queryIdx].pt)
                p2.append(kp[f.trainIdx].pt)
 
        # 單應(yīng)性矩陣
        H, _ = cv2.findHomography(np.float32(p2), np.float32(p1), cv2.RHO)
 
        # 透視變換
        output_frame = cv2.warpPerspective(frame, H, self.__capture['size'], borderMode=cv2.BORDER_REPLICATE)
 
        return output_frame
 
if __name__ == '__main__':
 
    if not os.path.exists(input_path):
        print(f'[ERROR] File "{input_path}" not found')
        exit(0)
    else:
        print(f'[INFO] Video "{input_path}" stable begin')
 
    s = Stable()
    s.stable(input_path, output_path, number)
 
    print('[INFO] Done.')
    exit(0)