Appearance
第十六章 · 光流与运动分析
16.1 光流（Optical Flow）
光流 = 视频中相邻帧之间像素的运动场。Lukas-Kanade 和 Farneback 是最常用的两种方法。
16.1.1 Lucas-Kanade 光流
适用于稀疏光流 — 跟踪特征点的运动。
python
import cv2
import numpy as np
# 准备两帧
prev_gray = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
next_gray = cv2.cvtColor(frame2, cv2.COLOR_BGR2GRAY)
# 用 goodFeaturesToTrack 找初始特征点
prev_pts = cv2.goodFeaturesToTrack(
prev_gray, maxCorners=100, qualityLevel=0.01, minDistance=20, blockSize=3
)
# ---- calcOpticalFlowPyrLK ----
# prevImg: 前一帧
# nextImg: 后一帧
# prevPts: 输入特征点 (N, 1, 2)
# nextPts: 输出特征点 (N, 1, 2)
# status: 1=找到, 0=丢失
# err: 跟踪误差
next_pts, status, err = cv2.calcOpticalFlowPyrLK(
prev_gray, next_gray,
prev_pts, None,
winSize=(31, 31), # 搜索窗口大小
maxLevel=3, # 金字塔层数
criteria=(cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_MAX_ITER, 40, 0.001)
)
# 筛选有效跟踪点
good_new = next_pts[status == 1]
good_old = prev_pts[status == 1]
# 可视化运动向量
img2 = next_gray.copy()
for i, (new, old) in enumerate(zip(good_new, good_old)):
a, b = new.ravel()
c, d = old.ravel()
cv2.line(img2, (int(a), int(b)), (int(c), int(d)), (0, 255, 0), 2)
cv2.circle(img2, (int(a), int(b)), 3, (0, 255, 0), -1)16.1.2 Farneback 稠密光流
计算每个像素的光流。
python
# ---- calcOpticalFlowFarneback ----
# prev: 前一帧灰度
# next: 后一帧灰度
# pyr_scale: 金字塔比例 (0.5=每层减半)
# levels: 金字塔层数
# winsize: 窗口大小
# iterations: 每层迭代次数
# poly_n: 多项式展开大小 (5-7 推荐)
# poly_sigma: 标准差
# flags: 0=默认, OPTFLOW_USE_INITIAL_FLOW=使用初始估计, OPTFLOW_LK_GET_MIN_EIGENVALS=最小特征值
prev_gray = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
next_gray = cv2.cvtColor(frame2, cv2.COLOR_BGR2GRAY)
flow = cv2.calcOpticalFlowFarneback(
prev_gray, next_gray,
None,
pyr_scale=0.5,
levels=3,
winsize=15,
iterations=3,
poly_n=5,
poly_sigma=1.2,
flags=0
)
# flow shape: (h, w, 2) — flow[y, x] = (u, v) = (dx, dy)16.1.3 光流可视化
python
def visualize_flow(flow):
"""可视化光流场"""
h, w = flow.shape[:2]
u = flow[:,:,0]
v = flow[:,:,1]
magnitude = np.sqrt(u**2 + v**2)
magnitude = magnitude / magnitude.max() * 255 # 归一化
angle = np.arctan2(v, u) * 180 / np.pi
# HSV → BGR 映射
hue = np.abs(60 * (angle + 90) / 180) % 180 # 角度 → 色相
value = magnitude.astype(np.uint8)
hsv = np.zeros((h, w, 3), dtype=np.uint8)
hsv[:,:,0] = hue.astype(np.uint8)
hsv[:,:,1] = 255
hsv[:,:,2] = value
bgr = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)
return bgr
flow_img = visualize_flow(flow)
cv2.imshow('Optical Flow', flow_img)
cv2.waitKey(0)16.1.4 光流方法对比
| 方法 | 类型 | 速度 | 精度 | 适用场景 |
|---|---|---|---|---|
| LK | 稀疏 | ⚡⚡⚡ | ★★★★ | 特征点跟踪 |
| Farneback | 稠密 | ⚡⚡ | ★★★★ | 整体运动场 |
| Dense | 稠密 | ⚡ | ★★ | 复杂运动 |
| DIS | 稠密 | ⚡⚡ | ★★ | 大位移 |
16.2 运动分析（背景减除 + 运动检测）
16.2.1 帧差法
python
# 简单帧差 — 相邻帧差分
gray1 = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
gray2 = cv2.cvtColor(frame2, cv2.COLOR_BGR2GRAY)
diff = cv2.absdiff(gray1, gray2)
_, motion = cv2.threshold(diff, 30, 255, cv2.THRESH_BINARY)
# 形态学清理
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
motion = cv2.morphologyEx(motion, cv2.MORPH_OPEN, kernel)
motion = cv2.morphologyEx(motion, cv2.MORPH_CLOSE, kernel)16.2.2 混合高斯背景建模（MOG2）
python
# ---- MOG2 背景减除 ----
fgmg = cv2.createBackgroundSubtractorMOG2(
history=500, # 缓冲区历史帧数
varThreshold=16, # 方差阈值（检测灵敏度）
detectShadows=True # 检测阴影
)
# 逐帧处理
while True:
fg_mask = fgmg.apply(frame)
# fg_mask: 前景掩码（255=前景, 0=背景, 128=阴影）
# 清理掩码
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
fg_mask = cv2.morphologyEx(fg_mask, cv2.MORPH_OPEN, kernel)
fg_mask = cv2.morphologyEx(fg_mask, cv2.MORPH_CLOSE, kernel)
if fgmg.detectShadows():
# 将阴影区域标记为 128，可用以下方法移除
fg_mask[fg_mask == 128] = 0
# 检测运动对象
contours, _ = cv2.findContours(fg_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
for cnt in contours:
if cv2.contourArea(cnt) > 1000: # 最小面积过滤
x, y, w, h = cv2.boundingRect(cnt)
cv2.rectangle(frame, (x, y), (x+w, y+h), (0, 255, 0), 2)16.2.3 KNN 背景建模
python
# ---- KNN 背景减除（更鲁棒）----
fgbg = cv2.createBackgroundSubtractorKNN(
history=500,
dist2Threshold=400.0, # 距离阈值
detectShadows=True
)
fg_mask = fgbg.apply(frame)16.2.4 运动检测完整管线
python
def motion_detect_pipeline(frame, fgbg, min_area=1000):
"""运动检测完整管线"""
fg_mask = fgbg.apply(frame)
# 清理
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
fg_mask = cv2.morphologyEx(fg_mask, cv2.MORPH_OPEN, kernel, iterations=2)
fg_mask = cv2.morphologyEx(fg_mask, cv2.MORPH_CLOSE, kernel, iterations=2)
# 连通域分析
num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(fg_mask, 8)
objects = []
for i in range(1, num_labels):
if stats[i, cv2.CC_STAT_AREA] >= min_area:
x, y, w, h, area = stats[i]
objects.append({'bbox': (x, y, w, h), 'area': area, 'center': centroids[i]})
return fg_mask, objects16.3 光流与运动分析速查表
| 方法 | 函数 | 输出 | 适用 |
|---|---|---|---|
| LK 稀疏光流 | calcOpticalFlowPyrLK() | 特征点位移 | 跟踪特定物体 |
| Farneback 稠密 | calcOpticalFlowFarneback() | (h,w,2) 光流场 | 整体运动分析 |
| 帧差 | cv2.absdiff() | 差分图 | 简单运动检测 |
| MOG2 | createBackgroundSubtractorMOG2() | 前景掩码 | 背景/前景分离 |
| KNN | createBackgroundSubtractorKNN() | 前景掩码 | 阴影检测 |