树莓派中使用线程调用OpenCV抓取图像帧为何变慢？

问题背景 #

在树莓派上使用Python+OpenCV读取RPi摄像头帧时，单循环模式下实时运行正常，但将摄像头操作移至后台线程按需取帧后，出现频繁不可预测的延迟，帧率降至每秒1-2帧。以下是原实现代码：

摄像头类代码 #

import cv2
import numpy as np
from threading import Thread, Lock

# 补充原代码缺失的cam_mode定义（示例）
class cam_mode:
    width = 640
    height = 480

class cam:
    def __init__(self):        
        self.cap = cv2.VideoCapture(0)
        self.frame = np.zeros([cam_mode.width, cam_mode.height, 3], dtype=np.uint8)
        
        self.cam_thread = Thread(target=self.camera_thread, args=())
        self._lock = Lock()
        # 补充原代码缺失的控制变量初始化
        self.cont = True
        self.is_running = True
        
    def start_camera(self):
        self.cam_thread.daemon = True
        self.cam_thread.start()
        
    def camera_thread(self):
        if not self.is_running:
            raise Exception("Camera is not running")
        while self.cont:
            with self._lock:
                status, frame = self.cap.read()                
                if status:
                    self.frame = frame
                    
    def get_image(self):
        with self._lock:
            return self.frame.copy() 

调用代码 #

import cv2

my_cam = cam()
my_cam.start_camera()
while True:
      image = my_cam.get_image()
      cv2.imshow('', image)
      if cv2.waitKey(1) == ord("q"):
          my_cam.cont = False
          break
cv2.destroyAllWindows()

核心问题与优化建议 #

• 缩小锁的作用范围（关键修复）
  原代码将耗时的cap.read()放入锁内，导致锁被长期占用，主线程取帧时频繁阻塞。应仅在更新共享的self.frame时加锁，取帧操作放在锁外：

  def camera_thread(self):
      if not self.is_running:
          raise Exception("Camera is not running")
      while self.cont:
          # 取帧操作不占用锁
          status, frame = self.cap.read()                
          if status:
              # 仅更新共享帧时加锁
              with self._lock:
                  self.frame = frame
  

• 移除不必要的帧拷贝
  若主线程仅用于显示帧、不修改帧内容，可直接返回帧引用（需在锁保护下读取），减少内存拷贝开销：

  def get_image(self):
      with self._lock:
          return self.frame  # 去掉copy()，前提是主线程不修改该帧
  

• 优化摄像头参数配置
  树莓派OpenCV默认参数未针对实时场景优化，手动设置分辨率、帧率和缓冲区大小，降低延迟：

  def __init__(self):        
      self.cap = cv2.VideoCapture(0)
      # 设置摄像头基础参数
      self.cap.set(cv2.CAP_PROP_FRAME_WIDTH, cam_mode.width)
      self.cap.set(cv2.CAP_PROP_FRAME_HEIGHT, cam_mode.height)
      self.cap.set(cv2.CAP_PROP_FPS, 30)
      # 减小缓冲区大小，避免帧堆积
      self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)
      # 初始化其他变量...
  

• 主线程帧率控制
  主线程无需无意义的频繁取帧，按目标帧率控制取帧频率，减少锁竞争：

  import time
  
  my_cam = cam()
  my_cam.start_camera()
  target_fps = 30
  frame_interval = 1.0 / target_fps
  last_frame_time = time.time()
  
  while True:
      current_time = time.time()
      if current_time - last_frame_time >= frame_interval:
          image = my_cam.get_image()
          cv2.imshow('', image)
          last_frame_time = current_time
      if cv2.waitKey(1) == ord("q"):
          my_cam.cont = False
          break
  cv2.destroyAllWindows()
  

内容的提问来源于stack exchange，提问作者Wiredchop