PYNQ-Z1实时人脸检测卡顿

Question

开发板PYNQ-Z1
代码如下，严重掉帧，会延迟大约三四秒，请问应该怎么优化？
用多进程或VDMA的优化方法要怎么实现？

import cv2
from pynq.overlays.base import BaseOverlay
from pynq.lib.video import *
print("base.bit")
base = BaseOverlay("base.bit")
# monitor configuration: 640*480 @ 60Hz
Mode = VideoMode(640, 480, 24)
hdmi_out = base.video.hdmi_out
hdmi_out.configure(Mode, PIXEL_BGR)
hdmi_out.start()

# camera (input) configuration
frame_in_w = 640
frame_in_h = 480

videoIn = cv2.VideoCapture(0)
videoIn.set(cv2.CAP_PROP_FRAME_WIDTH, frame_in_w);
videoIn.set(cv2.CAP_PROP_FRAME_HEIGHT, frame_in_h);
print("capture device is open: " + str(videoIn.isOpened()))
face_cascade = cv2.CascadeClassifier(
    '/home/xilinx/jupyter_notebooks/base/video/data/'
    'haarcascade_frontalface_default.xml')
eye_cascade = cv2.CascadeClassifier(
    '/home/xilinx/jupyter_notebooks/base/video/data/'
    'haarcascade_eye.xml')


while True:
    ret, frame_vga = videoIn.read()
    if ret==0:
        break
    np_frame = frame_vga
    gray = cv2.cvtColor(frame_vga, cv2.COLOR_BGR2GRAY)
    faces = face_cascade.detectMultiScale(gray, 1.3, 5)

    for (x,y,w,h) in faces:
        cv2.rectangle(np_frame,(x,y),(x+w,y+h),(255,0,0),2)
        roi_gray = gray[y:y+h, x:x+w]
        roi_color = np_frame[y:y+h, x:x+w]

        eyes = eye_cascade.detectMultiScale(roi_gray)
        for (ex,ey,ew,eh) in eyes:
            cv2.rectangle(roi_color,(ex,ey),(ex+ew,ey+eh),(0,255,0),2)
    
    if (ret):
        outframe = hdmi_out.newframe()
        outframe[:] = np_frame
        hdmi_out.writeframe(outframe)
    else:
        raise RuntimeError("Error while reading from camera.")

Answer 1

你可以尝试以下几种优化措施：

1. 调整视频帧率。降低帧率可以减少计算压力，从而提高性能。可以使用cv2.CAP_PROP_FPS属性来设置帧率，例如videoIn.set(cv2.CAP_PROP_FPS, 30)可以设置帧率为30fps。
2. 优化人脸检测算法。你可以尝试使用更高效的算法，例如MTCNN或RetinaFace，这些算法可以在保持准确性的同时提高性能。也可以尝试调整detectMultiScale方法的参数以获取更好的性能。
3. 降低视频分辨率。降低分辨率可以减少计算压力，从而提高性能。你可以尝试将frame_in_w和frame_in_h设置为更小的值，例如320x240或160x120。
4. 使用多线程或多进程。将视频处理和输出分别放在不同的线程或进程中，可以提高性能和稳定性。可以使用Python的multiprocessing或threading模块来实现。
   下面是优化后的代码示例：

import cv2
import time
from pynq.overlays.base import BaseOverlay
from pynq.lib.video import *
print("base.bit")
base = BaseOverlay("base.bit")
# monitor configuration: 640*480 @ 60Hz
Mode = VideoMode(640, 480, 24)
hdmi_out = base.video.hdmi_out
hdmi_out.configure(Mode, PIXEL_BGR)
hdmi_out.start()

# camera (input) configuration
frame_in_w = 640
frame_in_h = 480

videoIn = cv2.VideoCapture(0)
videoIn.set(cv2.CAP_PROP_FRAME_WIDTH, frame_in_w);
videoIn.set(cv2.CAP_PROP_FRAME_HEIGHT, frame_in_h);
print("capture device is open: " + str(videoIn.isOpened()))
face_cascade = cv2.CascadeClassifier(
    '/home/xilinx/jupyter_notebooks/base/video/data/'
    'haarcascade_frontalface_default.xml')
eye_cascade = cv2.CascadeClassifier(
    '/home/xilinx/jupyter_notebooks/base/video/data/'
    'haarcascade_eye.xml')

frame_rate = 10 # 每秒处理10帧视频
delay = 1/frame_rate

while True:
    ret, frame_vga = videoIn.read()
    if ret==0:
        break
    np_frame = frame_vga
    gray = cv2.cvtColor(frame_vga, cv2.COLOR_BGR2GRAY)

    # 只检测视频帧的中心区域
    x, y, w, h = int(frame_in_w/4), int(frame_in_h/4), int(frame_in_w/2), int(frame_in_h/2)
    gray = gray[y:y+h, x:x+w]
    faces = face_cascade.detectMultiScale(gray, 1.3, 5)

    for (x,y,w,h) in faces:
        cv2.rectangle(np_frame,(x+int(frame_in_w/4),y+int(frame_in_h/4)),
            (x+int(frame_in_w/4)+w,y+int(frame_in_h/4)+h),(255,0,0),2)
        roi_gray = gray[y:y+h, x:x+w]
        roi_color = np_frame[y+int(frame_in_h/4):y+int(frame_in_h/4)+h, 
            x+int(frame_in_w/4):x+int(frame_in_w/4)+w]

        eyes = eye_cascade.detectMultiScale