Python實現(xiàn)視頻目標檢測與軌跡跟蹤流程詳解
更新時間：2023年01月28日 09:38:00 作者：傾城一少
通過閱讀相關文獻及測試，找到了一種基于多模板匹配的改進方法，可以對遙感視頻衛(wèi)星中的移動目標進行探測，并繪制其軌跡。根據(jù)實驗結(jié)果發(fā)現(xiàn)，可以比較有效的對運動目標進行跟蹤
一、原理
核心思想比較簡單。即通過不同旋轉(zhuǎn)角度的模板同時匹配，在多個結(jié)果中，找到相似度最大的結(jié)果，即認為匹配成功。 在視頻的某一幀將這些模板分別進行匹配，即可獲得較為準確的結(jié)果。
某一幀的物體搜索窗口如上圖所示。0°表示提取的原始模板，將原始模板以8個方向進行旋轉(zhuǎn)，可得到8個不同旋轉(zhuǎn)角度的模板。 依次與窗口進行模板匹配，可以得到相似度。取相似度最大的模板對應的坐標結(jié)果作為軌跡。
同時根據(jù)不同的精度需求，可以有4模板、8模板和16模板，對應方向如下。模板數(shù)目越多，其對旋轉(zhuǎn)的檢測性就越好、越精確。但同時計算量也會成倍增加。
二、代碼實現(xiàn)
# 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
# ---------------必要參數(shù)---------------
# 待識別視頻路徑
video_path = 'E:\\object\\test_real.mp4'
# 衛(wèi)星視頻地表分辨率
resolution = 2
# 估計最快運動速度
velocity = 850
# ---------------必要參數(shù)---------------
# ---------------可選參數(shù)---------------
# 提取的模板是否為正方形
isSquare = True
# 是否自動根據(jù)速度信息計算閾值
isAutoDisThresh = True
# 是否為多模板
isMultiTemplate = True
# 是否采用均值對軌跡進行平滑
isSmooth = True
# 相鄰軌跡點之間的距離閾值
dis_thresh = 10
# 多模板個數(shù)
templateNum = 8
# 初始待選窗口大小半徑
range_d = 30
# 灰度閾值敏感度，越大灰度閾值越低
gray_factor = 0.2
# 識別框縮放因子，越大繪制的識別框越大
scale_factor = 1.5
# 模板縮放因子，越大模板圖像越大
template_factor = 0.6
# 識別框顏色
color = (0, 0, 255)
# 輸出路徑
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"
# ---------------可選參數(shù)---------------
# 循環(huán)變量
count = 0
# 打開視頻
cap = cv2.VideoCapture(video_path)
cap2 = cv2.VideoCapture(video_path)
# 獲取視頻圖像大小
# video_h對應豎直方向，video_w對應水平方向
video_h = int(cap.get(4))
video_w = int(cap.get(3))
total = int(cap.get(7))
# 新建一張與視頻等大的影像用于繪制軌跡
track = np.zeros((video_h, video_w, 3), np.uint8)
# tlp用于存放待選窗口的左上角點
tlp = []
# rbp用于存放待選窗口的右下角點
rbp = []
# bottom_right_points用于存放目標區(qū)域的右下角點
bottom_right_points = []
# center_points用于存放目標區(qū)域的中心點
center_points = []
# trackPoints用于存放目標區(qū)域的左上角點
trackPoints = []
# Vs用于存放目標各幀速度
Vs = []
# 根據(jù)視頻信息計算每一幀的等待時間
if cap.get(5) != 0:
waitTime = int(1000.0 / cap.get(5))
fps = cap.get(5)
# 如果為真，則自動確定距離閾值
if isAutoDisThresh:
# 計算物體幀間最大運動范圍(像素)
max_range = math.ceil((5.0 * velocity) / (18.0 * resolution * (fps - 1)))
# 計算最大移動距離，作為閾值
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))
# 首先提取模板圖像
if cap2.isOpened():
# 讀取前兩幀
ret, frame1 = cap2.read()
ret, frame2 = cap2.read()
# 相減做差
sub = cv2.subtract(frame1, frame2)
# 得到的結(jié)果灰度化
gray = cv2.cvtColor(sub, cv2.COLOR_BGR2GRAY)
# 判斷作差后的結(jié)果是否全為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為中心，range_d為距離向外拓展得到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
# 一些越界的判斷
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
# 根據(jù)窗口坐標提取窗口內(nèi)容
win_ini = cv2.cvtColor(frame1[win_tl_y:win_rb_y, win_tl_x:win_rb_x, :], cv2.COLOR_BGR2GRAY)
# 獲取最大值位置對應的灰度值
tem_img = cv2.cvtColor(frame1, cv2.COLOR_BGR2GRAY)
# 由最大值對應灰度值計算合適的灰度閾值
gray_thresh = tem_img[loc_y, loc_x] - gray_factor * tem_img[loc_y, loc_x]
# 初始窗口二值化處理
ret, thresh = cv2.threshold(win_ini, gray_thresh, 255, cv2.THRESH_BINARY)
# 在初始窗口中尋找輪廓
img2, contours, hi = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
# 有可能找到多個輪廓，但認為包含點數(shù)最多的那個輪廓是要找的輪廓
length = []
for item in contours:
length.append(item.shape[0])
target_contour = contours[length.index(max(length))]
# 獲取目標輪廓的坐標信息
x, y, w, h = cv2.boundingRect(target_contour)
if isSquare:
# 保證提取的模板為正方形
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:
# 不保證模板為正方形
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
# 一些越界判斷
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
# 提取模板內(nèi)容
template = frame1[real_tl_y:real_rb_y, real_tl_x:real_rb_x, :]
# 獲取模板的寬高，h豎直方向，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)
# 計算待選窗口左上角點坐標
tlx = loc_x - d
tly = loc_y - d
# 判斷是否越界，越界則設置為0
if tlx < 0:
tlx = 0
if tly < 0:
tly = 0
range_tl = (tlx, tly)
# 計算待選窗口右下角點坐標
rbx = loc_x + w + d
rby = loc_y + h + d
# 判斷是否越界，越界設置為視頻長寬最大值
if rbx > video_w:
rbx = video_w
if rby > video_h:
rby = video_h
range_rb = (rbx, rby)
# 放入角點坐標列表
tlp.append(range_tl)
rbp.append(range_rb)
cap2.release()
# 然后進行模板匹配
while cap.isOpened():
# 讀取每幀內(nèi)容
ret, frame = cap.read()
# 判斷幀內(nèi)容是否為空，不為空繼續(xù)
if frame is None:
break
else:
# 是否為多模板匹配模式
if isMultiTemplate:
if templateNum == 16:
# 逐個模板進行匹配
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)
# 獲取各模板對應的最大值
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)
# 尋找最佳匹配結(jié)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)
# 獲取最佳匹配結(jié)果對應的坐標信息
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坐標順序(水平，豎直)(→，↓)
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:
# 計算待選窗口左上角點坐標
tlx = top_left[0] - d
tly = top_left[1] - d
# 判斷是否越界，越界則設置為0
if tlx < 0:
tlx = 0
if tly < 0:
tly = 0
range_tl = (tlx, tly)
# 計算待選窗口右下角點坐標
rbx = top_left[0] + w + d
rby = top_left[1] + h + d
# 判斷是否越界，越界設置為視頻長寬最大值
if rbx > video_w:
rbx = video_w
if rby > video_h:
rby = video_h
range_rb = (rbx, rby)
# 將待選窗口左上角點坐標和右下角點坐標依次添加到列表中
tlp.append(range_tl)
rbp.append(range_rb)
# 將目標區(qū)域的左上角點、中心點、右下角點坐標依次加入列表
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:
# 加入運動連續(xù)性約束，若相鄰軌跡點距離相差大于閾值，則認為錯誤
distance = abs(trackPoints[-1][0] - top_left[0]) + abs(trackPoints[-1][1] - top_left[1])
if distance > dis_thresh:
print '100%'
break
else:
# 計算待選窗口左上角點坐標
tlx = top_left[0] - d
tly = top_left[1] - d
# 判斷是否越界，越界則設置為0
if tlx < 0:
tlx = 0
if tly < 0:
tly = 0
range_tl = (tlx, tly)
# 計算待選窗口右下角點坐標
rbx = top_left[0] + w + d
rby = top_left[1] + h + d
# 判斷是否越界，越界設置為視頻長寬最大值
if rbx > video_w:
rbx = video_w
if rby > video_h:
rby = video_h
range_rb = (rbx, rby)
# 將待選窗口左上角點坐標和右下角點坐標依次添加到列表中
tlp.append(range_tl)
rbp.append(range_rb)
# 將目標區(qū)域的左上角點、中心點、右下角點坐標依次加入列表
trackPoints.append(top_left)
bottom_right_points.append(bottom_right)
# 判斷是否采用均值平滑
if isSmooth:
# 采用均值平滑，平滑軌跡
center_point = ((center_point[0] + center_points[-1][0]) / 2,
(center_point[1] + center_points[-1][1]) / 2)
center_points.append(center_point)
# 繪制目標識別框
cv2.rectangle(frame,
(center_point[0] - offset, center_point[1] - offset),
(center_point[0] + offset, center_point[1] + offset),
color, 2)
# 繪制運動軌跡
cv2.line(track, center_points[-2], center_points[-1], (255, 255, 255), 1)
# 計算速度
Vs.append(calcVelocity(center_points[-2][0],
center_points[-1][0],
center_points[-2][1],
center_points[-1][1],
resolution,
waitTime))
# 輸出目標、軌跡視頻
out.write(frame)
out2.write(track)
count += 1
print round((count * 1.0 / total) * 100, 2), '%'
# 顯示結(jié)果
cv2.imshow("Tr", track)
cv2.imshow("Fr", frame)
# 退出控制
k = cv2.waitKey(waitTime) & 0xFF
if k == 27:
break
# 打印軌跡坐標
print trackPoints
print '相鄰幀距離閾值:', dis_thresh
print '灰度閾值:', gray_thresh
print '模板縮放因子:', template_factor
print '識別框縮放因子:', scale_factor
# 輸出中心點軌跡
output = open(out_path3, 'w')
for item in center_points:
output.write(item.__str__() + "\n")
# 輸出各幀速度
output2 = open(out_path4, 'w')
for item in Vs:
output2.write(item.__str__() + "\n")
# 釋放對象
cap.release()
out.release()
out2.release()
output.close()
output2.close()在代碼中主要做了如下改進：
1.增加多模板匹配機制
為了能精確地檢測物體的旋轉(zhuǎn)，引入多模板匹配。在代碼中有4、8、16不同數(shù)量的模式可選。模板越多，對于旋轉(zhuǎn)的識別越精確。 下圖匹配模板數(shù)分別是1、4、8、16。
可以看到，單模版匹配已經(jīng)無法正常識別跟蹤了。模板數(shù)為4時，會有少量跟蹤錯誤。當模板數(shù)為8和16時，跟蹤的軌跡就相對精確了。 下圖是采用8模板和單模板匹配的軌跡比較，可以看到，利用多模板匹配，可以較好識別旋轉(zhuǎn)物體。 白色為單模版匹配軌跡，紅色為多模板匹配軌跡。
同時考慮到衛(wèi)星視頻動目標一般運動形式是平移和旋轉(zhuǎn)，沒有縮放。所以經(jīng)過優(yōu)化的算法可以滿足大部分需求。
2.增加軌跡平滑
通過對軌跡列表中最后兩個點求均值作為最終的軌跡點，可以對提取的軌跡進行一定程度的平滑。
三、測試對比
下圖是模擬飛機曲線飛行的視頻。對其進行目標識別和軌跡提取后如下。
對應的飛行軌跡如下。
可以看到，相較于單模版匹配，能較好地提取運動目標和軌跡。而采用之前的單模版匹配算法，經(jīng)過測試在剛轉(zhuǎn)彎時就跟丟了，如下。
到此這篇關于Python實現(xiàn)視頻目標檢測與軌跡跟蹤流程詳解的文章就介紹到這了,更多相關Python視頻目標檢測內(nèi)容請搜索腳本之家以前的文章或繼續(xù)瀏覽下面的相關文章希望大家以后多多支持腳本之家！
相關文章
這篇文章主要介紹了詳解python里的命名規(guī)范，命名應當盡量使用全拼寫的單詞，縮寫的情況文章中也給大家提到，需要的朋友參考下吧2018-07-07
這篇文章主要為大家詳細介紹了Python中使用支持向量機SVM實踐，具有一定的參考價值，感興趣的小伙伴們可以參考一下2017-12-12
python GUI庫圖形界面開發(fā)之PyQt5輸入對話框QInputDialog詳細使用方法與實例
這篇文章主要介紹了python GUI庫圖形界面開發(fā)之PyQt5輸入對話框QInputDialog詳細使用方法與實例,需要的朋友可以參考下2020-02-02
因項目需要最近接手維護后續(xù)python程序并做后續(xù)開發(fā),下面這篇文章主要給大家介紹了關于Python?3.x踩坑的相關資料,文中通過實例代碼介紹的非常詳細,需要的朋友可以參考下2022-03-03
Python用yield from實現(xiàn)異步協(xié)程爬蟲的實踐
本文主要介紹了Python用yield from實現(xiàn)異步協(xié)程爬蟲的實踐，文中通過示例代碼介紹的非常詳細，對大家的學習或者工作具有一定的參考學習價值，需要的朋友們下面隨著小編來一起學習學習吧2023-01-01
這篇文章主要為大家介紹了Python?NumPy科學計算庫的高級應用深入詳解，有需要的朋友可以借鑒參考下，希望能夠有所幫助，祝大家多多進步，早日升職加薪2023-07-07
這篇文章主要介紹了Python實現(xiàn)微信翻譯機器人的方法，本文通過實例代碼給大家介紹的非常詳細，具有一定的參考借鑒價值,需要的朋友可以參考下2019-08-08
這篇文章主要介紹了python實現(xiàn)單向鏈表詳解，分享了相關代碼示例，每一步操作前都有簡單分析，小編覺得還是挺不錯的，具有一定借鑒價值，需要的朋友可以參考下2018-02-02
這篇文章主要為大家詳細介紹了如何利用python實現(xiàn)電腦自動關機，具有一定的參考價值，感興趣的小伙伴們可以參考一下2018-06-06
利用Python統(tǒng)計Jira數(shù)據(jù)并可視化
目前公司使用 Jira 作為項目管理工具，在每一次迭代完成后的復盤會上，我們都需要針對本次迭代的 Bug 進行數(shù)據(jù)統(tǒng)計，以幫助管理層能更直觀的了解研發(fā)的代碼質(zhì)量。本篇文章將介紹如何利用統(tǒng)計 Jira 數(shù)據(jù)，并進行可視化，需要的可以參考一下2022-07-07