def extract_feature(img):
hog_features = features.hog_feature(img,
orient=orient,
pix_per_cell=pix_per_cell,
cell_per_block=cell_per_block,
hog_channel=hog_channel)
spatial_features = features.bin_spatial(img, size=spatial_size)
hist_features = features.color_hist(img, nbins=hist_bins)
return np.hstack((spatial_features, hist_features, hog_features))
def single_img_features(self, img):
#1) Define an empty list to receive features
img_features = []
#2) Apply color conversion if other than 'RGB'
if self.color_space != 'RGB':
if self.color_space == 'HSV':
feature_image = cv2.cvtColor(img, cv2.COLOR_RGB2HSV)
elif self.color_space == 'LUV':
feature_image = cv2.cvtColor(img, cv2.COLOR_RGB2LUV)
elif self.color_space == 'HLS':
feature_image = cv2.cvtColor(img, cv2.COLOR_RGB2HLS)
elif self.color_space == 'YUV':
feature_image = cv2.cvtColor(img, cv2.COLOR_RGB2YUV)
elif self.color_space == 'YCrCb':
feature_image = cv2.cvtColor(img, cv2.COLOR_RGB2YCrCb)
else:
feature_image = np.copy(img)
#3) Compute spatial features if flag is set
if self.spatial_feat == True:
spatial_features = features.bin_spatial(feature_image,
size=self.spatial_size)
#4) Append features to list
img_features.append(spatial_features)
#5) Compute histogram features if flag is set
if self.hist_feat == True:
hist_features = features.color_hist(feature_image,
nbins=self.hist_bins)
#6) Append features to list
img_features.append(hist_features)
#7) Compute HOG features if flag is set
if self.hog_feat == True:
if self.hog_channel == 'ALL':
hog_features = []
for channel in range(feature_image.shape[2]):
hog_features.extend(
features.get_hog_features(feature_image[:, :, channel],
self.orient,
self.pix_per_cell,
self.cell_per_block,
vis=False,
feature_vec=True))
else:
hog_features = features.get_hog_features(
feature_image[:, :, self.hog_channel],
self.orient,
self.pix_per_cell,
self.cell_per_block,
vis=False,
feature_vec=True)
#8) Append features to list
img_features.append(hog_features)
#9) Return concatenated array of features
return np.concatenate(img_features)
def find_cars(img, scale, ystart, ystop, pix_per_cell, cell_per_block, orient, spatial_size, hist_bins):
draw_img = np.copy(img)
# Make a heatmap of zeros
heatmap = np.zeros_like(img[:,:,0])
img = img.astype(np.float32)/255
img_tosearch = img[ystart:ystop,:,:]
ctrans_tosearch = convert_color(img_tosearch)
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(ctrans_tosearch, (np.int(imshape[1]/scale), np.int(imshape[0]/scale)))
ch1 = ctrans_tosearch[:,:,0]
ch2 = ctrans_tosearch[:,:,1]
ch3 = ctrans_tosearch[:,:,2]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - 1
nyblocks = (ch1.shape[0] // pix_per_cell) - 1
nfeat_per_block = orient*cell_per_block**2
window = 64
nblocks_per_window = (window // pix_per_cell) - 1
cells_per_step = 2
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step
nysteps = (nyblocks - nblocks_per_window) // cells_per_step
# Compute individual channel HOG features for the entire image
hog1 = get_hog_features(ch1, orient, pix_per_cell, cell_per_block, feature_vec=False)
hog2 = get_hog_features(ch2, orient, pix_per_cell, cell_per_block, feature_vec=False)
hog3 = get_hog_features(ch3, orient, pix_per_cell, cell_per_block, feature_vec=False)
for xb in range(nxsteps):
for yb in range(nysteps):
ypos = yb*cells_per_step
xpos = xb*cells_per_step
hog_feat1 = hog1[ypos:ypos+nblocks_per_window, xpos:xpos+nblocks_per_window].ravel()
hog_feat2 = hog2[ypos:ypos+nblocks_per_window, xpos:xpos+nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos+nblocks_per_window, xpos:xpos+nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
xleft = xpos*pix_per_cell
ytop = ypos*pix_per_cell
# Extract the image patch
subimg = cv2.resize(ctrans_tosearch[ytop:ytop+window, xleft:xleft+window], (64, 64))
# Get color features
spatial_features = bin_spatial(subimg, size=spatial_size)
hist_features = color_hist(subimg, nbins=hist_bins)
# Scale features and make a prediction
test_features = X_scaler.transform(np.hstack((spatial_features, hist_features, hog_features)).reshape(1, -1))
test_prediction = svc.predict(test_features)
if test_prediction == 1:
xbox_left = np.int(xleft*scale)
ytop_draw = np.int(ytop*scale)
win_draw = np.int(window*scale)
cv2.rectangle(draw_img, (xbox_left, ytop_draw+ystart),
(xbox_left+win_draw, ytop_draw+win_draw+ystart), (0, 0, 255), 6)
heatmap[ytop_draw+ystart:ytop_draw+win_draw+ystart, xbox_left:xbox_left+win_draw] += 1
return draw_img, heatmap
def find_cars(img,
ystart,
ystop,
scale,
svc,
X_scaler,
orient=9,
pix_per_cell=8,
cell_per_block=2,
spatial_size=(32, 32),
hist_bins=32,
spatial_feat=True,
hist_feat=True,
hog_feat=True,
vis_boxes=False):
# draw_img = np.copy(img)
boxes = []
spatial_features = []
hog_features = []
img = img.astype(np.float32) / 255
img_tosearch = img[ystart:ystop, :, :]
ctrans_tosearch = convert_color(img_tosearch, conv='RGB2YCrCb')
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(
ctrans_tosearch,
(np.int(imshape[1] / scale), np.int(imshape[0] / scale)))
ch1 = ctrans_tosearch[:, :, 0]
ch2 = ctrans_tosearch[:, :, 1]
ch3 = ctrans_tosearch[:, :, 2]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - cell_per_block + 1
nyblocks = (ch1.shape[0] // pix_per_cell) - cell_per_block + 1
# 64 was the orginal sampling rate, with 8 cells and 8 pix per cell
window = 64
nblocks_per_window = (window // pix_per_cell) - cell_per_block + 1
cells_per_step = 2 # Instead of overlap, define how many cells to step
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step + 1
nysteps = (nyblocks - nblocks_per_window) // cells_per_step + 1
if hog_feat:
# Compute individual channel HOG features for the entire image
hog1 = get_hog_features(ch1,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog2 = get_hog_features(ch2,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog3 = get_hog_features(ch3,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
for xb in range(nxsteps):
for yb in range(nysteps):
ypos = yb * cells_per_step
xpos = xb * cells_per_step
# Extract HOG for this patch
if hog_feat:
hog_feat1 = hog1[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat2 = hog2[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
xleft = xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
# Get color features
if spatial_feat:
spatial_features = bin_spatial(subimg, size=spatial_size)
if hist_feat:
hist_features = color_hist(subimg, nbins=hist_bins)
# Scale features and make a prediction
test_features = X_scaler.transform(
np.hstack((spatial_features, hist_features,
hog_features)).reshape(1, -1))
# test_features = X_scaler.transform(np.hstack((shape_feat, hist_feat)).reshape(1, -1))
test_prediction = svc.predict(test_features)
if test_prediction == 1 or vis_boxes:
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
startx = xbox_left
starty = ytop_draw + ystart
endx = xbox_left + win_draw
endy = ytop_draw + win_draw + ystart
boxes.append(((startx, starty), (endx, endy)))
return boxes
def img_process_pipeline_2(images,
orient,
pix_per_cell,
cell_per_block,
spatial_size,
hist_bins,
scaler,
svc,
saveFig=False):
"""
Simplified / faster image processing pipeline
:param images:
:return:
"""
out_images = []
out_maps = []
out_titles = []
out_boxes = []
ystart = 400
ystop = 656
scale = 1.5
# Iterate over test images
for img_src in images:
img_boxes = []
t = time.time()
count = 0
# Read in the image
img = mpimg.imread(img_src)
draw_img = np.copy(img)
# Make a heatmap of zeros
heatmap = np.zeros_like(img[:, :, 0])
img = img.astype(np.float32) / 255
img_tosearch = img[ystart:ystop, :, :]
ctrans_tosearch = convert_color(img_tosearch, conv='RGB2YCrCb')
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(
ctrans_tosearch,
(np.int(imshape[1] / scale), np.int(imshape[0] / scale)))
ch1 = ctrans_tosearch[:, :, 0]
ch2 = ctrans_tosearch[:, :, 1]
ch3 = ctrans_tosearch[:, :, 2]
# Define blocks and steps
nxblocks = (ch1.shape[1] // pix_per_cell) - 1
nyblocks = (ch1.shape[0] // pix_per_cell) - 1
nfeat_per_block = orient * cell_per_block**2
window = 64
nblocks_per_window = (window // pix_per_cell) - 1
cells_per_step = 2
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step
nysteps = (nyblocks - nblocks_per_window) // cells_per_step
# Compute individual channel HOG features for the entire image
hog1 = get_hog_features(ch1,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog2 = get_hog_features(ch2,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog3 = get_hog_features(ch3,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
for xb in range(nxsteps):
for yb in range(nysteps):
count += 1
xpos = xb * cells_per_step
ypos = yb * cells_per_step
# Extract HOG for this patch
hog_feat1 = hog1[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat2 = hog2[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
xleft = xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
# Get color features
spatial_features = bin_spatial(subimg, size=spatial_size)
hist_features = color_hist(subimg, nbins=hist_bins)
# Scale features and make a prediction
test_features = scaler.transform(
np.hstack((spatial_features, hist_features,
hog_features)).reshape(1, -1))
test_pred = svc.predict(test_features)
if test_pred == 1:
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
# Draw a box
cv2.rectangle(
draw_img, (xbox_left, ytop_draw + ystart),
(xbox_left + win_draw, ytop_draw + win_draw + ystart),
(0, 0, 255))
img_boxes.append(((xbox_left, ytop_draw + ystart),
(xbox_left + win_draw,
ytop_draw + win_draw + ystart)))
# Draw a heatmap
heatmap[ytop_draw + ystart:ytop_draw + win_draw + ystart,
xbox_left:xbox_left + win_draw] += 1
print(time.time() - t, 'seconds to run, total windows =', count)
out_images.append(draw_img)
out_titles.append(img_src[-9:])
out_titles.append(img_src[-9:])
out_images.append(heatmap)
out_maps.append(heatmap)
out_boxes.append(img_boxes)
fig = plt.figure(figsize=(12, 24))
visualize(fig, 8, 2, out_images, out_titles)
plt.savefig('output_images/streamlined.png')
def find_cars(self, img):
t1 = time.time()
ystart = 400
ystop = 656
scale = 1
orient = 9
pix_per_cell = 8
cell_per_block = 2
spatial_size = (32, 32)
hist_bins = 32
draw_img = np.copy(img)
img = img.astype(np.float32) / 255
img_tosearch = img[ystart:ystop, :, :]
ctrans_tosearch = self.convert_color(img_tosearch, conv='RGB2YCrCb')
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(
ctrans_tosearch,
(np.int(imshape[1] / scale), np.int(imshape[0] / scale)))
# visualizer.draw_image(ctrans_tosearch)
ch1 = ctrans_tosearch[:, :, 0]
ch2 = ctrans_tosearch[:, :, 1]
ch3 = ctrans_tosearch[:, :, 2]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - cell_per_block + 1
nyblocks = (ch1.shape[0] // pix_per_cell) - cell_per_block + 1
nfeat_per_block = orient * cell_per_block**2
# 64 was the orginal sampling rate, with 8 cells and 8 pix per cell
window = 64
nblocks_per_window = (window // pix_per_cell) - cell_per_block + 1
cells_per_step = 2 # Instead of overlap, define how many cells to step
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step
nysteps = (nyblocks - nblocks_per_window) // cells_per_step
# Compute individual channel HOG features for the entire image
hog1 = features.get_hog_features(ch1,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog2 = features.get_hog_features(ch2,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog3 = features.get_hog_features(ch3,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
for xb in range(nxsteps):
for yb in range(nysteps):
ypos = yb * cells_per_step
xpos = xb * cells_per_step
# Extract HOG for this patch
hog_feat1 = hog1[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat2 = hog2[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
xleft = xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
# Get color features
spatial_features = features.bin_spatial(subimg,
size=spatial_size)
hist_features = features.color_hist(subimg, nbins=hist_bins)
# Scale features and make a prediction
test_features = self.X_scaler.transform(
np.hstack((spatial_features, hist_features,
hog_features)).reshape(1, -1))
# test_features = X_scaler.transform(np.hstack((shape_feat, hist_feat)).reshape(1, -1))
test_prediction = self.svc.predict(test_features)
if test_prediction == 1:
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
cv2.rectangle(
draw_img, (xbox_left, ytop_draw + ystart),
(xbox_left + win_draw, ytop_draw + win_draw + ystart),
(0, 0, 255), 6)
t2 = time.time()
print(round(t2 - t1, 2), 'Seconds to classify one image')
return draw_img
def find_cars(img, scaler, model, config):
orient = config["orient"]
pix_per_cell = config["pix_per_cell"]
cell_per_block = config["cell_per_block"]
spatial_size = config["spatial_size"]
hist_bins = config["hist_bins"]
train_image_format = config["train_image_format"]
colorspace = config["colorspace"]
hog_channel = config["hog_channel"]
spatial_feat = config["spatial_feat"]
hist_feat = config["hist_feat"]
hog_feat = config["hog_feat"]
ystart = config["y_start"]
ystop = config["y_stop"]
scale = config["scale"]
boxes = []
# if you extracted training
# data from .png images (scaled 0 to 1 by mpimg) and the
# image you are searching is a .jpg (scaled 0 to 255)
if train_image_format == 'png':
img = img.astype(np.float32) / 255
img_tosearch = img[ystart:ystop, :, :]
ctrans_tosearch = features.convert_color(img_tosearch, conv=colorspace)
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(
ctrans_tosearch,
(np.int(imshape[1] / scale), np.int(imshape[0] / scale)))
if hog_channel == 'ALL':
ch1 = ctrans_tosearch[:, :, 0]
ch2 = ctrans_tosearch[:, :, 1]
ch3 = ctrans_tosearch[:, :, 2]
else:
ch1 = ctrans_tosearch[:, :, hog_channel]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - cell_per_block + 1
nyblocks = (ch1.shape[0] // pix_per_cell) - cell_per_block + 1
nfeat_per_block = orient * cell_per_block**2
# 64 was the orginal sampling rate, with 8 cells and 8 pix per cell
window = 64
nblocks_per_window = (window // pix_per_cell) - cell_per_block + 1
cells_per_step = 2 # Instead of overlap, define how many cells to step
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step + 1
nysteps = (nyblocks - nblocks_per_window) // cells_per_step + 1
# Compute individual channel HOG features for the entire image
if hog_channel == 'ALL':
hog1 = features.get_hog_features(ch1,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog2 = features.get_hog_features(ch2,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog3 = features.get_hog_features(ch3,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
else:
hog1 = features.get_hog_features(ch1,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
# import time
#
# start = time.time()
# print("nxsteps=",nxsteps,"nysteps=",nysteps )
for xb in range(nxsteps):
for yb in range(nysteps):
# start_i = time.time()
ypos = yb * cells_per_step
xpos = xb * cells_per_step
# Extract HOG for this patch
if hog_channel == 'ALL':
hog_feat1 = hog1[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat2 = hog2[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
else:
hog_feat1 = hog1[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1))
# print("xb=",xb,"yb=",yb , "hog = ", time.time() - start_i)
xleft = xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
# Get color features
spatial_features = features.bin_spatial(subimg, size=spatial_size)
hist_features = features.color_hist(subimg, nbins=hist_bins)
# print("xb=", xb, "yb=", yb, "spatial/hist = ", time.time() - start_i)
# Scale features and make a prediction
test_features = scaler.transform(
np.hstack((spatial_features, hist_features,
hog_features)).reshape(1, -1))
# print("xb=", xb, "yb=", yb, "scale = ", time.time() - start_i)
# test_features = X_scaler.transform(np.hstack((shape_feat, hist_feat)).reshape(1, -1))
test_prediction = model.predict(test_features)
# print("xb=", xb, "yb=", yb, "pred = ", time.time() - start_i)
# print(nxsteps, nysteps, test_prediction)
if test_prediction == 1:
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
box = ((xbox_left, ytop_draw + ystart),
(xbox_left + win_draw, ytop_draw + win_draw + ystart))
boxes.append(box)
# print("xb=", xb, "yb=", yb, "total = ", time.time() - start_i)
# end = time.time()
# print("pred + sliding window = ", end - start)
return boxes
def find_cars(
self,
img,
y_start_stop,
window=64,
cells_per_step=2 # Instead of overlap, define how many cells to step
):
ystart, ystop = y_start_stop
ystart = ystart or 0
ystop = ystop or img.shape[0]
ctrans_tosearch = img[ystart:ystop, :, :]
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(
ctrans_tosearch,
(np.int(imshape[1] / scale), np.int(imshape[0] / scale)))
ch1 = ctrans_tosearch[:, :, 0]
ch2 = ctrans_tosearch[:, :, 1]
ch3 = ctrans_tosearch[:, :, 2]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - 1
nyblocks = (ch1.shape[0] // pix_per_cell) - 1
# nfeat_per_block = orient * cell_per_block**2
# 64 was the orginal sampling rate, with 8 cells and 8 pix per cell
nblocks_per_window = (window // pix_per_cell) - 1
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step
nysteps = (nyblocks - nblocks_per_window) // cells_per_step
# Compute individual channel HOG features for the entire image
hog1 = features.get_hog_features(ch1,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog2 = features.get_hog_features(ch2,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog3 = features.get_hog_features(ch3,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
for xb in range(nxsteps):
for yb in range(nysteps):
ypos = yb * cells_per_step
xpos = xb * cells_per_step
# Extract HOG for this patch
hog_feat1 = hog1[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat2 = hog2[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
xleft = xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
# Get color features
spatial_features = features.bin_spatial(subimg,
size=spatial_size)
hist_features = features.color_hist(subimg, nbins=hist_bins)
# Scale features and make a prediction
test_features = self.X_scaler.transform(
np.hstack((spatial_features, hist_features,
hog_features)).reshape(1, -1))
test_prediction = self.clf.predict(test_features)
if test_prediction == 1:
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
yield xbox_left, ytop_draw + ystart, xbox_left + win_draw, ytop_draw + win_draw + ystart
def find_cars(img, ystart, ystop, scale, svc, X_scaler, orient, pix_per_cell,
cell_per_block, spatial_size, hist_bins, spatial_feat, hist_feat,
hog_feat, hog_channel):
draw_img = np.copy(img)
img = img.astype(np.float32) / 255
img_tosearch = img[ystart:ystop, :, :]
ctrans_tosearch = convert_color(img_tosearch, color_space='YCrCb')
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(
ctrans_tosearch,
(np.int(imshape[1] / scale), np.int(imshape[0] / scale)))
ch1 = ctrans_tosearch[:, :, 0]
ch2 = ctrans_tosearch[:, :, 1]
ch3 = ctrans_tosearch[:, :, 2]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - cell_per_block + 1
nyblocks = (ch1.shape[0] // pix_per_cell) - cell_per_block + 1
nfeat_per_block = orient * cell_per_block**2
# 64 was the orginal sampling rate, with 8 cells and 8 pix per cell
window = 64
nblocks_per_window = (window // pix_per_cell) - cell_per_block + 1
cells_per_step = 2 # Instead of overlap, define how many cells to step
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step
nysteps = (nyblocks - nblocks_per_window) // cells_per_step
if hog_channel == 'ALL':
img_hog_features = []
for channel in range(ctrans_tosearch.shape[2]):
img_hog_features.append(
get_hog_features(ctrans_tosearch[:, :, channel],
orient,
pix_per_cell,
cell_per_block,
feature_vec=False))
else:
img_hog_features = get_hog_features(ctrans_tosearch[:, :, hog_channel],
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
bboxes = []
debug_boxes = []
for xb in range(nxsteps):
for yb in range(nysteps):
ypos = yb * cells_per_step
xpos = xb * cells_per_step
xleft = xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
#1) Define an empty list to receive features
img_features = []
#3) Compute spatial features if flag is set
if spatial_feat == True:
spatial_features = bin_spatial(subimg, size=spatial_size)
#4) Append features to list
img_features.append(spatial_features)
#5) Compute histogram features if flag is set
if hist_feat == True:
hist_features = color_hist(subimg, nbins=hist_bins)
#6) Append features to list
img_features.append(hist_features)
#7) Compute HOG features if flag is set
if hog_feat == True:
if hog_channel == 'ALL':
hog_features = []
for channel in range(ctrans_tosearch.shape[2]):
hog_features.extend(img_hog_features[channel][
ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel())
else:
hog_features = img_hog_features[
ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
#8) Append features to list
img_features.append(hog_features)
# Scale features and make a prediction
test_features = X_scaler.transform(
np.concatenate(img_features).reshape(1, -1))
test_prediction = svc.predict(test_features)
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
box = ((xbox_left, ytop_draw + ystart),
(xbox_left + win_draw, ytop_draw + win_draw + ystart))
debug_boxes.append(box)
if test_prediction == 1:
bboxes.append(box)
return bboxes, debug_boxes
def find_cars(img,
xstart,
xstop,
ystart,
ystop,
scale,
clf,
scaler,
cspace,
orient,
pix_per_cell,
cell_per_block,
spatial_size,
hist_bins,
log=None,
min_conf=0.4):
"""
A function that find cars in a frame.
It takes a region of interest of an image, converts it to a given color space, and runs a sliding window search
through the image searching for matches by extracting a series of features and feeding them to the provided
(trained) classifier, returning a list of found matches.
The features computed are: HOG (using sub-sampling), spatial binning and color histogram.
:param img: The input image
:param xstart: Minimum X coordinate defining the region of interest
:param xstop: Maximum X coordinate defining the region of interest
:param ystart: Minimum Y coordinate defining the region of interest
:param ystop: Maximum Y coordinate defining the region of interest
:param scale: Defines the size of the window (scale * (64, 64))
:param clf: The trained classifier
:param scaler: A trained scaler used to normalize the features
:param cspace: The target color space for the image before extracting features
:param orient: The number of HOG orientations bins for the histogram
:param pix_per_cell: 2-tuple specifying the size of each cell for extracting HOG
:param cell_per_block: 2-tuple defining the area (in cells) over which normalization is performed during HOG
:param spatial_size: 2-tuple defining the size for spatial binning
:param hist_bins: The number of bins for each channel of color histogram
:param log: A dictionary to log how many windows were searched and how many found cars
:param min_conf: The minimum prediction confidence score to approve a prediction
:return: A list of bounding boxes for every positive prediction from the classifier
"""
img_tosearch = img[ystart:ystop, xstart:xstop, :]
ctrans_tosearch = convert_color(img_tosearch, cspace)
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(
ctrans_tosearch,
(np.int(imshape[1] / scale), np.int(imshape[0] / scale)))
ch1 = ctrans_tosearch[:, :, 0]
ch2 = ctrans_tosearch[:, :, 1]
ch3 = ctrans_tosearch[:, :, 2]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - cell_per_block + 1
nyblocks = (ch1.shape[0] // pix_per_cell) - cell_per_block + 1
# 64 was the original sampling rate, with 8 cells and 8 pix per cell
window = 64
nblocks_per_window = (window // pix_per_cell) - cell_per_block + 1
# With the definition below, an overlap of 75% is defined
cells_per_step = 1 # Instead of overlap, define how many cells to step
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step + 1
nysteps = (nyblocks - nblocks_per_window) // cells_per_step + 1
# Compute individual channel HOG features for the entire image
hog1 = get_hog_features(ch1,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog2 = get_hog_features(ch2,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
hog3 = get_hog_features(ch3,
orient,
pix_per_cell,
cell_per_block,
feature_vec=False)
window_list = []
for xb in range(nxsteps):
for yb in range(nysteps):
ypos = yb * cells_per_step
xpos = xb * cells_per_step
# Extract HOG for this patch
hog_feat1 = hog1[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat2 = hog2[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
xleft = xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
# Get color features
spatial_features = bin_spatial(subimg, spatial_size)
hist_features = color_hist(subimg, hist_bins)
# Scale features and make a prediction
test_features = scaler.transform(
np.hstack((spatial_features, hist_features,
hog_features)).reshape(1, -1))
test_prediction = clf.predict(test_features)
conf = clf.decision_function(test_features)
# Record prediction
if log is not None:
if scale not in log:
log[scale] = [0, 0]
log[scale][1] += 1
if test_prediction == 1 and conf > min_conf:
log[scale][0] += 1
if test_prediction == 1 and conf > min_conf:
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
window_list.append(((xbox_left + xstart, ytop_draw + ystart),
(xbox_left + win_draw + xstart,
ytop_draw + win_draw + ystart)))
return window_list
def find_cars(img, ystart, ystop, scale, svc, X_scaler, orient, pix_per_cell,
cell_per_block, spatial_size, hist_bins):
draw_img = np.copy(img)
img = img.astype(np.float32) / 255
img_tosearch = img[ystart:ystop, :, :] # sub-sampling
ctrans_tosearch = convert_color(img_tosearch, conv='RGB2YUV')
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(
ctrans_tosearch,
(np.int(imshape[1] / scale), np.int(imshape[0] / scale)))
ch1 = ctrans_tosearch[:, :, 0]
ch2 = ctrans_tosearch[:, :, 1]
ch3 = ctrans_tosearch[:, :, 2]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - cell_per_block + 1
nyblocks = (ch1.shape[0] // pix_per_cell) - cell_per_block + 1
nfeat_per_block = orient * cell_per_block**2
# 64 was the orginal sampling rate, with 8 cells and 8 pix per cell
window = 64
nblocks_per_window = (window // pix_per_cell) - cell_per_block + 1
#nblocks_per_window = (window // pix_per_cell)-1
cells_per_step = 2 # Instead of overlap, define how many cells to step
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step
nysteps = (nyblocks - nblocks_per_window) // cells_per_step
# Compute individual channel HOG features for the entire image
hog1 = get_hog_features(ch1,
orient,
pix_per_cell,
cell_per_block,
vis=False,
feature_vec=False)
hog2 = get_hog_features(ch2,
orient,
pix_per_cell,
cell_per_block,
vis=False,
feature_vec=False)
hog3 = get_hog_features(ch3,
orient,
pix_per_cell,
cell_per_block,
vis=False,
feature_vec=False)
bboxes = []
for xb in range(nxsteps):
for yb in range(nysteps):
ypos = yb * cells_per_step
xpos = xb * cells_per_step
# Extract HOG for this patch
hog_feat1 = hog1[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat2 = hog2[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
xleft = xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
# Get color features
spatial_features = bin_spatial(subimg, size=spatial_size)
hist_features = color_hist(subimg, nbins=hist_bins)
# Scale features and make a prediction
test_stacked = np.hstack((spatial_features, hist_features,
hog_features)).reshape(1, -1)
test_features = X_scaler.transform(test_stacked)
#test_features = scaler.transform(np.array(features).reshape(1, -1))
#test_features = X_scaler.transform(np.hstack((shape_feat, hist_feat)).reshape(1, -1))
test_prediction = svc.predict(test_features)
if test_prediction == 1:
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
cv2.rectangle(
draw_img, (xbox_left, ytop_draw + ystart),
(xbox_left + win_draw, ytop_draw + win_draw + ystart),
(0, 0, 255), 6)
bboxes.append(((int(xbox_left), int(ytop_draw + ystart)),
(int(xbox_left + win_draw),
int(ytop_draw + win_draw + ystart))))
return draw_img, bboxes
def find_cars(image, y_start_stop, scale, clf, X_scaler, color_space,
spatial_size, hist_bins, orient, pix_per_cell, cell_per_block,
channels):
# draw_img = np.copy(image)
img = image.astype(np.float32) / 255
ystart, ystop = y_start_stop
img_tosearch = img[ystart:ystop, :, :]
ctrans_tosearch = color_convertor(color_space)(img_tosearch)
imshape = ctrans_tosearch.shape
nx, ny = np.int(imshape[1] / scale), np.int(imshape[0] / scale)
ctrans_tosearch = cv2.resize(ctrans_tosearch, (nx, ny))
# Define blocks and steps as above
nxblocks = (ctrans_tosearch.shape[1] // pix_per_cell) - cell_per_block + 1
nyblocks = (ctrans_tosearch.shape[0] // pix_per_cell) - cell_per_block + 1
# nfeat_per_block = orient * cell_per_block**2
# 64 was the orginal sampling rate, with 8 cells and 8 pix per cell
window = 64
nblocks_per_window = (window // pix_per_cell) - cell_per_block + 1
cells_per_step = 2 # Instead of overlap, define how many cells to step
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step + 1
nysteps = (nyblocks - nblocks_per_window) // cells_per_step + 1
x_stub = np.mod(ctrans_tosearch.shape[1], pix_per_cell * cell_per_block)
# Compute HOG features for the entire image, subsample later
hog_features_search = []
for c in channels:
hog = get_hog_features(ctrans_tosearch[:, x_stub:, c],
orient,
pix_per_cell,
cell_per_block,
vis=False,
feature_vec=False)
hog_features_search.append(hog)
box_list = []
for xb, yb in product(range(nxsteps), range(nysteps)):
ypos = yb * cells_per_step
xpos = xb * cells_per_step
# Extract HOG for this patch
hog_feats = []
for hog in hog_features_search:
feats = hog[ypos:ypos + nblocks_per_window,
xpos:xpos + nblocks_per_window]
hog_feats.append(feats.ravel())
hog_features = np.hstack(hog_feats)
xleft = x_stub + xpos * pix_per_cell
ytop = ypos * pix_per_cell
# Extract the image patch
subimg = cv2.resize(
ctrans_tosearch[ytop:ytop + window, xleft:xleft + window],
(64, 64))
# Get color features
spatial_features = bin_spatial(subimg, size=spatial_size)
hist_features = color_hist(subimg, nbins=hist_bins)
# Scale features and make a prediction
features = np.hstack((spatial_features, hist_features, hog_features))
test_features = X_scaler.transform(features.reshape(1, -1))
# test_features = X_scaler.transform(
# np.hstack((spatial_features, hist_features)).reshape(1, -1))
test_prediction = clf.predict(test_features)
if test_prediction == 1:
xbox_left = np.int(xleft * scale)
ytop_draw = np.int(ytop * scale)
win_draw = np.int(window * scale)
box = ((xbox_left, ytop_draw + ystart),
(xbox_left + win_draw, ytop_draw + win_draw + ystart))
# cv2.rectangle(draw_img, box[0], box[1], (0, 0, 255), 6)
box_list.append(box)
return box_list