本文档说明

使用 OpenCV 内置的 Haar 级联分类器，实现 USB 摄像头实时人脸检测与框选，并进行性能调优。

前置准备

source ~/opencv-env/bin/activate

Haar 级联分类器简介

OpenCV 内置了多种预训练的 Haar 级联分类器模型，无需额外下载：

模型文件	检测目标
haarcascade_frontalface_default.xml	正面人脸
haarcascade_eye.xml	眼睛
haarcascade_smile.xml	微笑
haarcascade_upperbody.xml	上半身

模型文件位于 OpenCV 安装目录中，可通过代码自动获取路径。

基础人脸检测

import cv2
import os

os.environ.setdefault('DISPLAY', ':0')

DEVICE_INDEX = 73

def main():
    # 加载人脸检测模型
    cascade_path = cv2.data.haarcascades + 'haarcascade_frontalface_default.xml'
    face_cascade = cv2.CascadeClassifier(cascade_path)

    cap = cv2.VideoCapture(DEVICE_INDEX)
    cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'))
    cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)

    if not cap.isOpened():
        print("Error: Cannot open camera")
        return

    win = "Face Detection"
    cv2.namedWindow(win, cv2.WINDOW_NORMAL)
    cv2.setWindowProperty(win, cv2.WND_PROP_FULLSCREEN, cv2.WINDOW_FULLSCREEN)

    print("Press 'q' to quit")

    import time
    frame_count = 0
    start_time = time.time()

    while True:
        ret, frame = cap.read()
        if not ret:
            break

        frame_count += 1

        # 转灰度（Haar 检测需要灰度图）
        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)

        # 检测人脸
        faces = face_cascade.detectMultiScale(
            gray,
            scaleFactor=1.3,
            minNeighbors=5,
            minSize=(30, 30)
        )

        # 画框
        for (x, y, w, h) in faces:
            cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
            cv2.putText(frame, "Face", (x, y - 10),
                        cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 255, 0), 2)

        # 显示 FPS 和人脸数量
        elapsed = time.time() - start_time
        fps = frame_count / elapsed if elapsed > 0 else 0
        cv2.putText(frame, f"FPS: {fps:.1f} | Faces: {len(faces)}", (10, 30),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 255, 255), 2)

        cv2.imshow(win, frame)

        if cv2.waitKey(1) & 0xFF == ord('q'):
            break

    cap.release()
    cv2.destroyAllWindows()
    print(f"Avg FPS: {fps:.1f}")

if __name__ == "__main__":
    main()

运行：

DISPLAY=:0 ~/opencv-env/bin/python3 face_detect.py

性能调优

detectMultiScale 参数说明

参数	作用	调优建议
scaleFactor	图像缩放比例	越大越快但可能漏检，推荐 1.2~1.5
minNeighbors	最少邻居数	越大误检越少，推荐 4~6
minSize	最小检测尺寸	过滤太小的误检，推荐 (30,30)

优化版：缩小检测图 + 跳帧

import cv2
import time
import os

os.environ.setdefault('DISPLAY', ':0')

DEVICE_INDEX = 73
DETECT_SCALE = 0.5  # 缩小到 50% 做检测
SKIP_FRAMES = 2     # 每 N 帧检测一次

def main():
    cascade_path = cv2.data.haarcascades + 'haarcascade_frontalface_default.xml'
    face_cascade = cv2.CascadeClassifier(cascade_path)

    cap = cv2.VideoCapture(DEVICE_INDEX)
    cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'))
    cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)

    if not cap.isOpened():
        print("Error: Cannot open camera")
        return

    win = "Face Detection (Optimized)"
    cv2.namedWindow(win, cv2.WINDOW_NORMAL)
    cv2.setWindowProperty(win, cv2.WND_PROP_FULLSCREEN, cv2.WINDOW_FULLSCREEN)

    frame_count = 0
    start_time = time.time()
    faces = []

    while True:
        ret, frame = cap.read()
        if not ret:
            break

        frame_count += 1

        # 每 N 帧检测一次，其余帧复用上次结果
        if frame_count % SKIP_FRAMES == 0:
            small = cv2.resize(frame, None, fx=DETECT_SCALE, fy=DETECT_SCALE)
            gray = cv2.cvtColor(small, cv2.COLOR_BGR2GRAY)
            faces = face_cascade.detectMultiScale(gray, 1.3, 5, minSize=(20, 20))

        # 画框（坐标需要还原）
        for (x, y, w, h) in faces:
            sx = int(x / DETECT_SCALE)
            sy = int(y / DETECT_SCALE)
            sw = int(w / DETECT_SCALE)
            sh = int(h / DETECT_SCALE)
            cv2.rectangle(frame, (sx, sy), (sx + sw, sy + sh), (0, 255, 0), 2)

        elapsed = time.time() - start_time
        fps = frame_count / elapsed if elapsed > 0 else 0
        cv2.putText(frame, f"FPS: {fps:.1f} | Faces: {len(faces)}", (10, 30),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 255, 255), 2)

        cv2.imshow(win, frame)

        if cv2.waitKey(1) & 0xFF == ord('q'):
            break

    cap.release()
    cv2.destroyAllWindows()
    print(f"Optimized Avg FPS: {fps:.1f}")

if __name__ == "__main__":
    main()

性能对比

方案	分辨率	帧率
基础版（每帧检测）	1280×720	13.1fps
优化版（缩小+跳帧）	640×480	19.3fps

进一步加速

如需更高性能的人脸检测/识别，可以使用 RK3576 的 NPU（RKNN）进行 AI 推理加速，推理速度可达 4.4ms/帧。