def main():
## flags
isLoadDepthData = False
## sliding window
winSize = 5
winStep = 3
## load depth data and save to file
if isLoadDepthData:
loadDepthData()
## load dataset
trainActionList = loadTrainDataset()
## process for training
for trainAction in trainActionList:
ifile = open(trainAction.depthSequenceFile, 'r')
depthSequence = pickle.load(ifile)
ifile.close()
xoyImgList = []
xozImgList = []
yozImgList = []
# get xoy, xoz and yoz images from depth data
for depthData in depthSequence:
# # depth image visualization
# # color map image
# grayImg = mat2gray(depthData)
# colorMapImg = applyColorMap(grayImg, cv2.COLORMAP_JET)
# cv2.imshow('', colorMapImg)
# cv2.waitKey()
# # point cloud visualization
# points = getWorldCoordinates(depthData)
# visualizePointCloud(points)
xoyImg = mat2gray(depthData)
xozImg, yozImg = getProjectionImages(depthData)
xoyImgList.append(xoyImg)
xozImgList.append(xozImg)
yozImgList.append(yozImg)
trainAction.xoyImgList = xoyImgList
trainAction.xozImgList = xozImgList
trainAction.yozImgList = yozImgList
# windowed DMHI
xoyDmhiList = calWinDepthMHIList(trainAction.xoyImgList, winSize, winStep)
xozDmhiList = calWinDepthMHIList(trainAction.xozImgList, winSize, winStep)
yozDmhiList = calWinDepthMHIList(trainAction.yozImgList, winSize, winStep)
# post processing, crop and resize
postXoyDmhiList = cropAndResizeImageList(xoyDmhiList)
postXozDmhiList = cropAndResizeImageList(xozDmhiList)
postYozDmhiList = cropAndResizeImageList(yozDmhiList)
def getDepthProjection(points, isCrop = True):
"""
FUNC: get projections by counting the number of 3D points
PARAM:
points: point clouds
crop: crop the ROI or not
RETURN:
resultProjImg: the projection image (range: 0-255)
"""
tmp_points = points.copy()
min_xw = np.min(tmp_points[:,0])
min_yw = np.min(tmp_points[:,1])
tmp_points[:,0] -= min_xw
tmp_points[:,1] -= min_yw
# width = math.ceil(np.max(tmp_points[:,0])) + 1
# height = math.ceil(np.max(tmp_points[:,1])) + 1
width = 800
height = 800
n_points, _ = tmp_points.shape
projData = np.zeros((height, width))
for n in xrange(n_points):
[xw, yw, zw] = tmp_points[n, :]
if zw != 0:
rowIdx = math.floor(yw)
colIdx = math.floor(xw)
projData[rowIdx, colIdx] += 1
projImg = mat2gray(projData) # scale to 0-255
# post processing
# closing (Dilation followed by Erosion)
kernel = np.ones((3, 3), np.uint8)
projImg1 = cv2.morphologyEx(projImg, cv2.MORPH_CLOSE, kernel)
projImg2 = cv2.equalizeHist(projImg1)
if isCrop:
boxRegion = findBoxRegion(projImg2)
top, bottom, left, right = boxRegion
resultProjImg = projImg2[top:bottom, left:right]
else:
resultProjImg = projImg2.copy()
return resultProjImg
def getDepthProjection(points, isCrop=True):
"""
FUNC: get projections by counting the number of 3D points
PARAM:
points: point clouds
crop: crop the ROI or not
RETURN:
resultProjImg: the projection image (range: 0-255)
"""
tmp_points = points.copy()
min_xw = np.min(tmp_points[:, 0])
min_yw = np.min(tmp_points[:, 1])
tmp_points[:, 0] -= min_xw
tmp_points[:, 1] -= min_yw
# width = math.ceil(np.max(tmp_points[:,0])) + 1
# height = math.ceil(np.max(tmp_points[:,1])) + 1
width = 800
height = 800
n_points, _ = tmp_points.shape
projData = np.zeros((height, width))
for n in xrange(n_points):
[xw, yw, zw] = tmp_points[n, :]
if zw != 0:
rowIdx = math.floor(yw)
colIdx = math.floor(xw)
projData[rowIdx, colIdx] += 1
projImg = mat2gray(projData) # scale to 0-255
# post processing
# closing (Dilation followed by Erosion)
kernel = np.ones((3, 3), np.uint8)
projImg1 = cv2.morphologyEx(projImg, cv2.MORPH_CLOSE, kernel)
projImg2 = cv2.equalizeHist(projImg1)
if isCrop:
boxRegion = findBoxRegion(projImg2)
top, bottom, left, right = boxRegion
resultProjImg = projImg2[top:bottom, left:right]
else:
resultProjImg = projImg2.copy()
return resultProjImg
def showDepthData(depthData, isColorMap = True):
"""
FUNC: show depth data
PARAM:
depthData: raw depth data
RETURN:
"""
depthImage = mat2gray(depthData)
if isColorMap:
depthImage = cv2.applyColorMap(depthImage, cv2.COLORMAP_JET)
cv2.imshow('', depthImage)
cv2.waitKey()
def showDepthData(depthData, isColorMap=True):
"""
FUNC: show depth data
PARAM:
depthData: raw depth data
RETURN:
"""
depthImage = mat2gray(depthData)
if isColorMap:
depthImage = cv2.applyColorMap(depthImage, cv2.COLORMAP_JET)
cv2.imshow('', depthImage)
cv2.waitKey()
def main():
## flags
isLoadDepthData = False
## sliding window
winSize = 5
winStep = 3
## load depth data and save to file
if isLoadDepthData:
loadDepthData()
## load dataset
trainActionList = loadTrainDataset()
## process for training
for trainAction in trainActionList:
ifile = open(trainAction.depthSequenceFile, 'r')
depthSequence = pickle.load(ifile)
ifile.close()
xoyImgList = []
xozImgList = []
yozImgList = []
# get xoy, xoz and yoz images from depth data
for depthData in depthSequence:
# # depth image visualization
# # color map image
# grayImg = mat2gray(depthData)
# colorMapImg = applyColorMap(grayImg, cv2.COLORMAP_JET)
# cv2.imshow('', colorMapImg)
# cv2.waitKey()
# # point cloud visualization
# points = getWorldCoordinates(depthData)
# visualizePointCloud(points)
xoyImg = mat2gray(depthData)
xozImg, yozImg = getProjectionImages(depthData)
xoyImgList.append(xoyImg)
xozImgList.append(xozImg)
yozImgList.append(yozImg)
trainAction.xoyImgList = xoyImgList
trainAction.xozImgList = xozImgList
trainAction.yozImgList = yozImgList
# windowed DMHI
xoyDmhiList = calWinDepthMHIList(trainAction.xoyImgList, winSize,
winStep)
xozDmhiList = calWinDepthMHIList(trainAction.xozImgList, winSize,
winStep)
yozDmhiList = calWinDepthMHIList(trainAction.yozImgList, winSize,
winStep)
# post processing, crop and resize
postXoyDmhiList = cropAndResizeImageList(xoyDmhiList)
postXozDmhiList = cropAndResizeImageList(xozDmhiList)
postYozDmhiList = cropAndResizeImageList(yozDmhiList)
def calDepthMHI(frames, motion_thresh = 10, stride = 1, isCrop = True):
"""
FUNC: Calculate DMHI from given sequence (a list of frames)
PARAM:
frames: a list of frames
motion_thresh: threshold for detection of motion region
stride: stride of calculation of difference between frames
isCrop: crop ROI or not
RETURN:
dmhi: depth motion history image
"""
duration = len(frames)
firstFrm = frames[0]
firstFrm = cv2.GaussianBlur(firstFrm, (3, 3), 0) # Gaussian blur
height, width = firstFrm.shape
# a set of DMHIs
D_MHIs = []
D_MHIs.append(np.zeros((height, width)).astype(np.int32))
for i in xrange(1, duration, stride):
prevFrmIndex = i - stride
if (prevFrmIndex >= 0):
# current image
currFrm = frames[i]
currFrm = cv2.GaussianBlur(currFrm, (3, 3), 0) # Gaussian blur
# previous image
prevFrm = frames[prevFrmIndex]
prevFrm = cv2.GaussianBlur(prevFrm, (3, 3), 0) # Gaussian blur
# frame difference
motionImg = depthFrameDiff(currFrm, prevFrm, motion_thresh)
# DMHI
# if D == 1
DMHI = motionImg.copy()
DMHI[np.where(motionImg == 1)] = duration
# otherwise
tmp = np.maximum(0, D_MHIs[-1] - 1)
idx = np.where(motionImg != 1)
DMHI[idx] = tmp[idx]
D_MHIs.append(DMHI)
# save mhi image to file
# dmhiImg = mat2gray(DMHI)
# cv2.imwrite('mhi_%i.png' %i, dmhiImg)
# get the result
finalDMHI = D_MHIs[-1]
# convert to image
dmhi = mat2gray(finalDMHI)
dmhi = cv2.GaussianBlur(dmhi, (3, 3), 0) # Gaussian blur
# crop or not
if isCrop:
boxRegion = findBoxRegion(dmhi)
top, bottom, left, right = boxRegion
resultDmhi = dmhi[top:bottom, left:right]
else:
resultDmhi = dmhi.copy()
return resultDmhi
