在这篇博客的最后，提到基于模板匹配的方法对于目标的旋转检测效果不是很好。 因此通过阅读相关文献及测试，找到了一种基于多模板匹配的改进方法，可以比较有效的对运动目标进行跟踪。
一、思路与原理
核心思想比较简单。即通过不同旋转角度的模板同时匹配，在多个结果中，找到相似度最大的结果，即认为匹配成功。
在视频的某一帧将这些模板分别进行匹配，即可获得较为准确的结果。
某一帧的物体搜索窗口如上图所示。0°表示提取的原始模板，将原始模板以8个方向进行旋转，可得到8个不同旋转角度的模板。
依次与窗口进行模板匹配，可以得到相似度。取相似度最大的模板对应的坐标结果作为轨迹。
同时根据不同的精度需求，可以有4模板、8模板和16模板，对应方向如下。
模板数目越多，其对旋转的检测性就越好、越精确。但同时计算量也会成倍增加。
二、代码实现
# coding=utf-8
import cv2
import numpy as np
import math
def calcVelocity(x1, x2, y1, y2, res, wT):
dist = pow(pow(y1 - y2, 2) + pow(x1 - x2, 2), 0.5) * res
v = dist / (wT / 1000.0) * 3.6
return v
# ---------------必要参数---------------
# 待识别视频路径
video_path = 'E:\\object\\test_real.mp4'
# 卫星视频地表分辨率
resolution = 2
# 估计最快运动速度
velocity = 850
# ---------------必要参数---------------
# ---------------可选参数---------------
# 提取的模板是否为正方形
isSquare = True
# 是否自动根据速度信息计算阈值
isAutoDisThresh = True
# 是否为多模板
isMultiTemplate = True
# 是否采用均值对轨迹进行平滑
isSmooth = True
# 相邻轨迹点之间的距离阈值
dis_thresh = 10
# 多模板个数
templateNum = 8
# 初始待选窗口大小半径
range_d = 30
# 灰度阈值敏感度，越大灰度阈值越低
gray_factor = 0.2
# 识别框缩放因子，越大绘制的识别框越大
scale_factor = 1.5
# 模板缩放因子，越大模板图像越大
template_factor = 0.6
# 识别框颜色
color = (0, 0, 255)
# 输出路径
parent_path = video_path.replace(video_path.split("\\")[-1], '')
out_path = parent_path + "object.avi"
out_path2 = parent_path + "track.avi"
out_path3 = parent_path + "points.txt"
out_path4 = parent_path + "velocity.txt"
out_path5 = parent_path + "template.jpg"
# ---------------可选参数---------------
# 循环变量
count = 0
# 打开视频
cap = cv2.VideoCapture(video_path)
cap2 = cv2.VideoCapture(video_path)
# 获取视频图像大小
# video_h对应竖直方向，video_w对应水平方向
video_h = int(cap.get(4))
video_w = int(cap.get(3))
total = int(cap.get(7))
# 新建一张与视频等大的影像用于绘制轨迹
track = np.zeros((video_h, video_w, 3), np.uint8)
# tlp用于存放待选窗口的左上角点
tlp = []
# rbp用于存放待选窗口的右下角点
rbp = []
# bottom_right_points用于存放目标区域的右下角点
bottom_right_points = []
# center_points用于存放目标区域的中心点
center_points = []
# trackPoints用于存放目标区域的左上角点
trackPoints = []
# Vs用于存放目标各帧速度
Vs = []
# 根据视频信息计算每一帧的等待时间
if cap.get(5) != 0:
waitTime = int(1000.0 / cap.get(5))
fps = cap.get(5)
# 如果为真，则自动确定距离阈值
if isAutoDisThresh:
# 计算物体帧间最大运动范围(像素)
max_range = math.ceil((5.0 * velocity) / (18.0 * resolution * (fps - 1)))
# 计算最大移动距离，作为阈值
dis_thresh = math.ceil(pow(pow(max_range, 2) + pow(max_range, 2), 0.5))
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter(out_path, fourcc, fps, (video_w, video_h))
out2 = cv2.VideoWriter(out_path2, fourcc, fps, (video_w, video_h))
# 首先提取模板图像
if cap2.isOpened():
# 读取前两帧
ret, frame1 = cap2.read()
ret, frame2 = cap2.read()
# 相减做差
sub = cv2.subtract(frame1, frame2)
# 得到的结果灰度化
gray = cv2.cvtColor(sub, cv2.COLOR_BGR2GRAY)
# 判断作差后的结果是否全为0
if gray.max() != 0:
# 找到最大值位置
loc = np.where(gray == gray.max())
loc_x = loc[1][0]
loc_y = loc[0][0]
# 以loc为中心，range_d为距离向外拓展得到window
win_tl_x = loc_x - range_d
win_tl_y = loc_y - range_d
win_rb_x = loc_x + range_d
win_rb_y = loc_y + range_d
# 一些越界的判断
if win_tl_x < 0:
win_tl_x = 0
if win_tl_y < 0:
win_tl_y = 0
if win_rb_x > video_w:
win_rb_x = video_w
if win_rb_y > video_h:
win_rb_y = video_h
# 根据窗口坐标提取窗口内容
win_ini = cv2.cvtColor(frame1[win_tl_y:win_rb_y, win_tl_x:win_rb_x, :], cv2.COLOR_BGR2GRAY)
# 获取最大值位置对应的灰度值
tem_img = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
# 由最大值对应灰度值计算合适的灰度阈值
gray_thresh = tem_img[loc_y, loc_x] - gray_factor * tem_img[loc_y, loc_x]
# 初始窗口二值化处理
ret, thresh = cv2.threshold(win_ini, gray_thresh, 255, cv2.THRESH_BINARY)
# 在初始窗口中寻找轮廓
img2, contours, hi = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
# 有可能找到多个轮廓，但认为包含点数最多的那个轮廓是要找的轮廓
length = []
for item in contours:
length.append(item.shape[0])
target_contour = contours[length.index(max(length))]
# 获取目标轮廓的坐标信息
x, y, w, h = cv2.boundingRect(target_contour)
if isSquare:
# 保证提取的模板为正方形
tem_tl_x = win_tl_x + x
tem_tl_y = win_tl_y + y
tem_rb_x = win_tl_x + x + w
tem_rb_y = win_tl_y + y + h
center_x = (tem_tl_x + tem_rb_x) / 2
center_y = (tem_tl_y + tem_rb_y) / 2
delta = int(template_factor * max(w, h))
real_tl_x = center_x - delta
real_rb_x = center_x + delta
real_tl_y = center_y - delta
real_rb_y = center_y + delta
else:
# 不保证模板为正方形
real_tl_x = win_tl_x + x
real_tl_y = win_tl_y + y
real_rb_x = win_tl_x + x + w
real_rb_y = win_tl_y + y + h
# 一些越界判断
if real_tl_x < 0:
real_tl_x = 0
if real_tl_y < 0:
real_tl_y = 0
if real_rb_x > video_w:
real_rb_x = video_w
if real_rb_y > video_h:
real_rb_y = video_h
# 提取模板内容
template = frame1[real_tl_y:real_rb_y, real_tl_x:real_rb_x, :]
# 获取模板的宽高，h竖直方向，w水平方向
h = template.shape[0]
w = template.shape[1]
d = max(w, h)
# 是否是多模板匹配
if isMultiTemplate:
if templateNum == 16:
M22_5 = cv2.getRotationMatrix2D((d / 2, d / 2), -22.5, 1)
M45 = cv2.getRotationMatrix2D((d / 2, d / 2), -45, 1)
M67_5 = cv2.getRotationMatrix2D((d / 2, d / 2), -67.5, 1)
M90 = cv2.getRotationMatrix2D((d / 2, d / 2), -90, 1)
M112_5 = cv2.getRotationMatrix2D((d / 2, d / 2), -112.5, 1)
M135 = cv2.getRotationMatrix2D((d / 2, d / 2), -135, 1)
M157_5 = cv2.getRotationMatrix2D((d / 2, d / 2), -157.5, 1)
M180 = cv2.getRotationMatrix2D((d / 2, d / 2), -180, 1)
M202_5 = cv2.getRotationMatrix2D((d / 2, d / 2), -202.5, 1)
M225 = cv2.getRotationMatrix2D((d / 2, d / 2), -225, 1)
M247_5 = cv2.getRotationMatrix2D((d / 2, d / 2), -247.5, 1)
M270 = cv2.getRotationMatrix2D((d / 2, d / 2), -270, 1)
M292_5 = cv2.getRotationMatrix2D((d / 2, d / 2), -292.5, 1)
M315 = cv2.getRotationMatrix2D((d / 2, d / 2), -315, 1)
M337_5 = cv2.getRotationMatrix2D((d / 2, d / 2), -337.5, 1)
template22_5 = cv2.warpAffine(template, M22_5, (d, d))
template45 = cv2.warpAffine(template, M45, (d, d))
template67_5 = cv2.warpAffine(template, M67_5, (d, d))
template90 = cv2.warpAffine(template, M90, (d, d))
template112_5 = cv2.warpAffine(template, M112_5, (d, d))
template135 = cv2.warpAffine(template, M135, (d, d))
template157_5 = cv2.warpAffine(template, M157_5, (d, d))
template180 = cv2.warpAffine(template, M180, (d, d))
template202_5 = cv2.warpAffine(template, M202_5, (d, d))
template225 = cv2.warpAffine(template, M225, (d, d))
template247_5 = cv2.warpAffine(template, M247_5, (d, d))
template270 = cv2.warpAffine(template, M270, (d, d))
template292_5 = cv2.warpAffine(template, M292_5, (d, d))
template315 = cv2.warpAffine(template, M315, (d, d))
template337_5 = cv2.warpAffine(template, M337_5, (d, d))
elif templateNum == 8:
M45 = cv2.getRotationMatrix2D((d / 2, d / 2), -45, 1)
M90 = cv2.getRotationMatrix2D((d / 2, d / 2), -90, 1)
M135 = cv2.getRotationMatrix2D((d / 2, d / 2), -135, 1)
M180 = cv2.getRotationMatrix2D((d / 2, d / 2), -180, 1)
M225 = cv2.getRotationMatrix2D((d / 2, d / 2), -225, 1)
M270 = cv2.getRotationMatrix2D((d / 2, d / 2), -270, 1)
M315 = cv2.getRotationMatrix2D((d / 2, d / 2), -315, 1)
template45 = cv2.warpAffine(template, M45, (d, d))
template90 = cv2.warpAffine(template, M90, (d, d))
template135 = cv2.warpAffine(template, M135, (d, d))
template180 = cv2.warpAffine(template, M180, (d, d))
template225 = cv2.warpAffine(template, M225, (d, d))
template270 = cv2.warpAffine(template, M270, (d, d))
template315 = cv2.warpAffine(template, M315, (d, d))
elif templateNum == 4:
M90 = cv2.getRotationMatrix2D((d / 2, d / 2), -90, 1)
M180 = cv2.getRotationMatrix2D((d / 2, d / 2), -180, 1)
M270 = cv2.getRotationMatrix2D((d / 2, d / 2), -270, 1)
template90 = cv2.warpAffine(template, M90, (d, d))
template180 = cv2.warpAffine(template, M180, (d, d))
template270 = cv2.warpAffine(template, M270, (d, d))
cv2.imshow("Template", template)
cv2.imwrite(out_path5, template)
offset = int(scale_factor * d)
# 计算待选窗口左上角点坐标
tlx = loc_x - d
tly = loc_y - d
# 判断是否越界，越界则设置为0
if tlx < 0:
tlx = 0
if tly < 0:
tly = 0
range_tl = (tlx, tly)
# 计算待选窗口右下角点坐标
rbx = loc_x + w + d
rby = loc_y + h + d
# 判断是否越界，越界设置为视频长宽最大值
if rbx > video_w:
rbx = video_w
if rby > video_h:
rby = video_h
range_rb = (rbx, rby)
# 放入角点坐标列表
tlp.append(range_tl)
rbp.append(range_rb)
cap2.release()
# 然后进行模板匹配
while cap.isOpened():
# 读取每帧内容
ret, frame = cap.read()
# 判断帧内容是否为空，不为空继续
if frame is None:
break
else:
# 是否为多模板匹配模式
if isMultiTemplate:
if templateNum == 16:
# 逐个模板进行匹配
res = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :], template,
cv2.TM_CCOEFF_NORMED)
res22_5 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template22_5,
cv2.TM_CCOEFF_NORMED)
res67_5 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template67_5,
cv2.TM_CCOEFF_NORMED)
res112_5 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template112_5,
cv2.TM_CCOEFF_NORMED)
res157_5 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template157_5,
cv2.TM_CCOEFF_NORMED)
res202_5 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template202_5,
cv2.TM_CCOEFF_NORMED)
res247_5 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template247_5,
cv2.TM_CCOEFF_NORMED)
res292_5 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template292_5,
cv2.TM_CCOEFF_NORMED)
res337_5 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template337_5,
cv2.TM_CCOEFF_NORMED)
res90 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template90,
cv2.TM_CCOEFF_NORMED)
res180 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template180,
cv2.TM_CCOEFF_NORMED)
res270 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template270,
cv2.TM_CCOEFF_NORMED)
res45 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template45,
cv2.TM_CCOEFF_NORMED)
res135 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template135,
cv2.TM_CCOEFF_NORMED)
res225 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template225,
cv2.TM_CCOEFF_NORMED)
res315 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template315,
cv2.TM_CCOEFF_NORMED)
# 获取各模板对应的最大值
m22_5 = np.max(res22_5)
m67_5 = np.max(res67_5)
m112_5 = np.max(res112_5)
m157_5 = np.max(res157_5)
m202_5 = np.max(res202_5)
m247_5 = np.max(res247_5)
m292_5 = np.max(res292_5)
m337_5 = np.max(res337_5)
m45 = np.max(res45)
m135 = np.max(res135)
m225 = np.max(res225)
m315 = np.max(res315)
m0 = np.max(res)
m90 = np.max(res90)
m180 = np.max(res180)
m270 = np.max(res270)
# 寻找最佳匹配结果
m = max(m0, m22_5, m45, m67_5, m90,
m112_5, m135, m157_5, m180,
m202_5, m225, m247_5, m270,
m292_5, m315, m337_5)
# 获取最佳匹配结果对应的坐标信息
if m == m0:
mIndex = 0
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res)
elif m == m90:
mIndex = 90
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res90)
elif m == m180:
mIndex = 180
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res180)
elif m == m270:
mIndex = 270
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res270)
elif m == m45:
mIndex = 45
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res45)
elif m == m135:
mIndex = 135
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res135)
elif m == m225:
mIndex = 225
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res225)
elif m == m315:
mIndex = 315
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res315)
elif m == m22_5:
mIndex = 22.5
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res22_5)
elif m == m67_5:
mIndex = 67.5
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res67_5)
elif m == m112_5:
mIndex = 112.5
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res112_5)
elif m == m157_5:
mIndex = 157.5
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res157_5)
elif m == m202_5:
mIndex = 202.5
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res202_5)
elif m == m247_5:
mIndex = 247.5
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res247_5)
elif m == m292_5:
mIndex = 292.5
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res292_5)
elif m == m337_5:
mIndex = 337.5
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res337_5)
elif templateNum == 8:
res = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :], template,
cv2.TM_CCOEFF_NORMED)
res90 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template90,
cv2.TM_CCOEFF_NORMED)
res180 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template180,
cv2.TM_CCOEFF_NORMED)
res270 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template270,
cv2.TM_CCOEFF_NORMED)
res45 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template45,
cv2.TM_CCOEFF_NORMED)
res135 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template135,
cv2.TM_CCOEFF_NORMED)
res225 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template225,
cv2.TM_CCOEFF_NORMED)
res315 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template315,
cv2.TM_CCOEFF_NORMED)
m45 = np.max(res45)
m135 = np.max(res135)
m225 = np.max(res225)
m315 = np.max(res315)
m0 = np.max(res)
m90 = np.max(res90)
m180 = np.max(res180)
m270 = np.max(res270)
m = max(m0, m45, m90, m135, m180, m225, m270, m315)
if m == m0:
mIndex = 0
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res)
elif m == m90:
mIndex = 90
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res90)
elif m == m180:
mIndex = 180
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res180)
elif m == m270:
mIndex = 270
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res270)
elif m == m45:
mIndex = 45
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res45)
elif m == m135:
mIndex = 135
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res135)
elif m == m225:
mIndex = 225
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res225)
elif m == m315:
mIndex = 315
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res315)
elif templateNum == 4:
res = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :], template,
cv2.TM_CCOEFF_NORMED)
res90 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template90,
cv2.TM_CCOEFF_NORMED)
res180 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template180,
cv2.TM_CCOEFF_NORMED)
res270 = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :],
template270,
cv2.TM_CCOEFF_NORMED)
m0 = np.max(res)
m90 = np.max(res90)
m180 = np.max(res180)
m270 = np.max(res270)
m = max(m0, m90, m180, m270)
if m == m0:
mIndex = 0
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res)
elif m == m90:
mIndex = 90
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res90)
elif m == m180:
mIndex = 180
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res180)
elif m == m270:
mIndex = 270
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res270)
else:
res = cv2.matchTemplate(frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :], template,
cv2.TM_CCOEFF_NORMED)
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res)
window = frame[tlp[count][1]:rbp[count][1], tlp[count][0]:rbp[count][0], :]
cv2.imshow("Window", window)
# top_left坐标顺序(水平，竖直)(→，↓)
top_left = (max_loc[0] + tlp[count][0], max_loc[1] + tlp[count][1])
bottom_right = (top_left[0] + w, top_left[1] + h)
center_point = ((top_left[0] + bottom_right[0]) / 2, (top_left[1] + bottom_right[1]) / 2)
if trackPoints.__len__() == 0:
# 计算待选窗口左上角点坐标
tlx = top_left[0] - d
tly = top_left[1] - d
# 判断是否越界，越界则设置为0
if tlx < 0:
tlx = 0
if tly < 0:
tly = 0
range_tl = (tlx, tly)
# 计算待选窗口右下角点坐标
rbx = top_left[0] + w + d
rby = top_left[1] + h + d
# 判断是否越界，越界设置为视频长宽最大值
if rbx > video_w:
rbx = video_w
if rby > video_h:
rby = video_h
range_rb = (rbx, rby)
# 将待选窗口左上角点坐标和右下角点坐标依次添加到列表中
tlp.append(range_tl)
rbp.append(range_rb)
# 将目标区域的左上角点、中心点、右下角点坐标依次加入列表
trackPoints.append(top_left)
bottom_right_points.append(bottom_right)
center_points.append(center_point)
cv2.circle(track, center_point, 2, (0, 0, 255), -1)
else:
# 加入运动连续性约束，若相邻轨迹点距离相差大于阈值，则认为错误
distance = abs(trackPoints[-1][0] - top_left[0]) + abs(trackPoints[-1][1] - top_left[1])
if distance > dis_thresh:
print '100%'
break
else:
# 计算待选窗口左上角点坐标
tlx = top_left[0] - d
tly = top_left[1] - d
# 判断是否越界，越界则设置为0
if tlx < 0:
tlx = 0
if tly < 0:
tly = 0
range_tl = (tlx, tly)
# 计算待选窗口右下角点坐标
rbx = top_left[0] + w + d
rby = top_left[1] + h + d
# 判断是否越界，越界设置为视频长宽最大值
if rbx > video_w:
rbx = video_w
if rby > video_h:
rby = video_h
range_rb = (rbx, rby)
# 将待选窗口左上角点坐标和右下角点坐标依次添加到列表中
tlp.append(range_tl)
rbp.append(range_rb)
# 将目标区域的左上角点、中心点、右下角点坐标依次加入列表
trackPoints.append(top_left)
bottom_right_points.append(bottom_right)
# 判断是否采用均值平滑
if isSmooth:
# 采用均值平滑，平滑轨迹
center_point = ((center_point[0] + center_points[-1][0]) / 2,
(center_point[1] + center_points[-1][1]) / 2)
center_points.append(center_point)
# 绘制目标识别框
cv2.rectangle(frame,
(center_point[0] - offset, center_point[1] - offset),
(center_point[0] + offset, center_point[1] + offset),
color, 2)
# 绘制运动轨迹
cv2.line(track, center_points[-2], center_points[-1], (255, 255, 255), 1)
# 计算速度
Vs.append(calcVelocity(center_points[-2][0],
center_points[-1][0],
center_points[-2][1],
center_points[-1][1],
resolution,
waitTime))
# 输出目标、轨迹视频
out.write(frame)
out2.write(track)
count += 1
print round((count * 1.0 / total) * 100, 2), '%'
# 显示结果
cv2.imshow("Tr", track)
cv2.imshow("Fr", frame)
# 退出控制
k = cv2.waitKey(waitTime) & 0xFF
if k == 27:
break
# 打印轨迹坐标
print trackPoints
print '相邻帧距离阈值:', dis_thresh
print '灰度阈值:', gray_thresh
print '模板缩放因子:', template_factor
print '识别框缩放因子:', scale_factor
# 输出中心点轨迹
output = open(out_path3, 'w')
for item in center_points:
output.write(item.__str__() + "\n")
# 输出各帧速度
output2 = open(out_path4, 'w')
for item in Vs:
output2.write(item.__str__() + "\n")
# 释放对象
cap.release()
out.release()
out2.release()
output.close()
output2.close()
在代码中主要做了如下改进。
1.增加多模板匹配机制
为了能精确地检测物体的旋转，引入多模板匹配。在代码中有4、8、16不同数量的模式可选。模板越多，对于旋转的识别越精确。
下图匹配模板数分别是1、4、8、16。
可以看到，单模版匹配已经无法正常识别跟踪了。模板数为4时，会有少量跟踪错误。当模板数为8和16时，跟踪的轨迹就相对精确了。
下图是采用8模板和单模板匹配的轨迹比较，可以看到，利用多模板匹配，可以较好识别旋转物体。
白色为单模版匹配轨迹，红色为多模板匹配轨迹。
同时考虑到卫星视频动目标一般运动形式是平移和旋转，没有缩放。所以经过优化的算法可以满足大部分需求。
2.增加轨迹平滑
通过对轨迹列表中最后两个点求均值作为最终的轨迹点，可以对提取的轨迹进行一定程度的平滑。如下图所示。
上图是未平滑轨迹，下图是平滑后轨迹。对于轨迹的某些突变点，均值平滑有很好的效果。
三、测试对比
下图是模拟飞机曲线飞行的视频。对其进行目标识别和轨迹提取后如下。
对应的飞行轨迹如下。
可以看到，相较于单模版匹配，能较好地提取运动目标和轨迹。而采用之前的单模版匹配算法，经过测试在刚转弯时就跟丢了，如下。
本文作者原创,未经许可不得转载,谢谢配合
