def Predict(self, img_path, vis_thresh=0.3, output_img="output.jpg"):
image = cv2.imread(img_path)
bboxes = self.system_dict["local"]["detector"](image)
image = draw_bboxes(image, bboxes, thresh=vis_thresh)
cv2.imwrite(output_img, image)
return bboxes
def main(args):
cam = cv2.VideoCapture(args.device)
if args.codec == 'YUY2':
cam.set(cv2.CAP_PROP_FOURCC, 844715353.0)
elif args.codec == 'MJPG':
cam.set(cv2.CAP_PROP_FOURCC, 0x47504A4D)
else:
print('use default video codec.')
if args.resolution:
cam.set(cv2.CAP_PROP_FRAME_WIDTH, args.resolution[0])
cam.set(cv2.CAP_PROP_FRAME_HEIGHT, args.resolution[1])
detector = CornerNet_Squeeze(
model_name=args.model) if args.model else CornerNet_Squeeze()
frame_count = 0
init_time = time()
tic = time()
try:
while True:
# Capture frame-by-frame
if cam.grab():
_, frame = cam.retrieve()
bboxes = detector(frame)
frame = draw_bboxes(frame, bboxes)
toc = time()
frame_count += 1
else:
continue
# Calculate fps
if toc - init_time > 3:
fps = frame_count / (toc - tic)
print('{:.2f}: {} x {} @ {:5.1f}'.format(
time(), frame.shape[1], frame.shape[0], fps))
if toc - tic > 3:
tic = time()
frame_count = 0
# Show the resulting frame
if args.visual:
frame = cv2.resize(frame, (0, 0), fx=args.scale, fy=args.scale)
cv2.imshow('/dev/video{}'.format(args.device), frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
except KeyboardInterrupt:
print('\nKeyboardInterrupt')
pass
# When everything done, release the capture
cam.release()
cv2.destroyAllWindows()
def detect_video(video_path, output_path, start=0, end=0, classes=None,
forbid_box=None):
detector = CornerNet_Saccade()
vid = cv2.VideoCapture(video_path)
if not vid.isOpened():
raise IOError("Couldn't open webcam or video")
video_fps = vid.get(cv2.CAP_PROP_FPS)
video_size = (int(vid.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT)))
print(output_path)
out_fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter(output_path, out_fourcc, video_fps, video_size)
if classes is not None:
classes = classes.split(',')
if forbid_box is not None:
forbid_box = [[[int(x) for x in b.split(',')]
for b in s.split(';')]
for s in forbid_box.split('|')]
start, end = int(start)*1000, int(end)
end = end if end == 0 else end*1000
width, height = 0, 0
while True:
return_value, frame = vid.read()
if return_value is False:
break
msec = int(vid.get(cv2.CAP_PROP_POS_MSEC))
if msec < start:
continue
if end > 0 and msec > end:
break
bboxes = detector(frame)
if colors is None:
set_colors([name for name in bboxes])
if classes is not None:
bboxes = {k: v for k, v in bboxes.items() if k in classes}
print('当前时间进度：%.2f秒' % (msec/1000))
image = draw_bboxes(frame, bboxes, colors=colors)
out.write(image)
print("width: %d, height: %d" % (width, height))
out.release()
def main(args):
cam = cv2.VideoCapture(args.filename)
detector = CornerNet_Squeeze(model_name=args.model) if args.model else CornerNet_Squeeze()
frame_count = 0
init_time = time()
tic = time()
try:
while True:
# Capture frame-by-frame
if cam.grab():
_, frame = cam.retrieve()
frame = cv2.resize(frame, (640,360))
bboxes = detector(frame)
frame = draw_bboxes(frame, bboxes)
toc = time()
frame_count += 1
else:
continue
# Calculate fps
if toc - init_time > 3:
fps = frame_count / (toc - tic)
print('{:.2f}: {} x {} @ {:5.1f}'.format(time(), frame.shape[1], frame.shape[0], fps))
if toc -tic > 3:
tic = time()
frame_count = 0
# Show the resulting frame
if args.visual:
frame = cv2.resize(frame, (0, 0), fx=args.scale, fy=args.scale)
cv2.imshow(args.filename, frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
except KeyboardInterrupt:
print('\nKeyboardInterrupt')
pass
# When everything done, release the capture
cam.release()
cv2.destroyAllWindows()
def Predict(self, img_path, vis_thresh=0.3, output_img="output.jpg"):
'''
User function: Run inference on multiple images and visualize them
Args:
img_path (str): Relative path to the image file
vis_thresh (float): Threshold for predicted scores. Scores for objects detected below this score will not be displayed
output_folder (str): Path to folder where output images will be saved
Returns:
None
'''
image = cv2.imread(img_path)
bboxes = self.system_dict["local"]["detector"](image)
image = draw_bboxes(image, bboxes, thresh=vis_thresh)
cv2.imwrite(output_img, image)
return bboxes
#!/usr/bin/env python
import cv2
from core.detectors import CornerNet_Saccade
# from core.detectors import CornerNet_Squeeze
from core.vis_utils import draw_bboxes
from core.paths import get_file_path
import os
import pickle
import pprint
detector = CornerNet_Saccade()
image = cv2.imread("./demo.jpg")
bboxes = detector(image)
pprint.pprint(bboxes)
# 为了支持中文显示，对此做了修改
# 注意修改自己数据集的id2label字典
image = draw_bboxes(image, bboxes)
cv2.imwrite("./demo_out.jpg", image)
if not cap.isOpened():
print("Camera is not ready ")
exit()
# 加载模型
detector = CornerNet_Squeeze()
while True:
ret, frame = cap.read()
if not ret:
pass
#image = cv2.imread("demo.jpg")
t0 = time.time()
#bboxes = detector(image)
bboxes = detector(frame)
# 画框
image = draw_bboxes(frame, bboxes)
#cv2.imwrite("demo_out.jpg", image)
# 计算 FPS
fps = 1.0 / (time.time() - t0)
print("[FPS]: ", fps)
cv2.imshow("cornernet", image)
cv2.waitKey(1)
import cv2
import time
from core.detectors import CornerNet_Saccade
from core.detectors import CornerNet_Squeeze
from core.detectors import CornerNet
from core.vis_utils import draw_bboxes
# detector = CornerNet()
# detector1 = CornerNet_Saccade()
detector2 = CornerNet_Squeeze()
####
####
start = time.time()
image = cv2.imread("demo.jpg")
# bboxes = detector(image)
# bboxes1 = detector1(image)
bboxes2 = detector2(image)
end = time.time()
print("time:", end - start)
# image = draw_bboxes(image, bboxes)
# image1 = draw_bboxes(image, bboxes1)
image2 = draw_bboxes(image, bboxes2)
# cv2.imwrite("demo_out_corner.jpg", image)
# cv2.imwrite("demo_out_saccade.jpg", image1)
cv2.imwrite("demo_out_squeeze.jpg", image2)
def cam(arg, detector):
count = 1
save_flag = False
print(arg)
current_path = os.getcwd()
model_dirpath = current_path + "/model"
clf = load_model(model_dirpath)
if arg == "video":
#cap = cv2.VideoCapture('/home/gisen/Documents/rosbag/2019-07-09-15-25-21.avi')
cap = cv2.VideoCapture('/home/gisen/Documents/rosbag/out_short.mp4')
width = int(cap.get(3))
height = int(cap.get(4))
writer = record(width, height)
elif arg == "camera":
cap = camera_open()
cap.set(cv2.CAP_PROP_FPS, 60) # カメラFPSを60FPSに設定
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640) # カメラ画像の横幅を1280に設定
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480) # カメラ画像の縦幅を720に設定
elif arg == "make_color4train":
cap = camera_open()
cap.set(cv2.CAP_PROP_FPS, 60) # カメラFPSを60FPSに設定
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640) # カメラ画像の横幅を1280に設定
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480) # カメラ画像の縦幅を720に設定
data_path = "/home/gisen/Documents/own_dataset/traffic_light_dataset/traffic_light/*"
imgs_path = load_color4train(data_path)
save_flag = True
#while True:
while (cap.isOpened()):
if arg == "video" or arg == "camera":
# VideoCaptureから1フレーム読み込む
ret, frame = cap.read()
if not ret:
print("画像の取得に失敗しました。")
continue
image, bboxes, bboxes_traffic, bboxes_pdstrn = obj_inference(
detector, frame)
traffic_trm_imges = []
pdstrn_trm_imges = []
trm_imges_dict = {}
bboxes_dict = {
"traffic_signal": bboxes_traffic,
"pedestrian_signal": bboxes_pdstrn
}
result = {}
if bboxes_traffic.shape[0] > 0:
try:
for bbox in bboxes_traffic:
x1 = int(bbox[0])
y1 = int(bbox[1])
x2 = int(bbox[2])
y2 = int(bbox[3])
trm_img = image[y1:y2, x1:x2]
traffic_trm_imges.append([trm_img])
except:
print("交通信号機のトリミングを試みましたが失敗しました")
if bboxes_pdstrn.shape[0] > 0:
try:
for bbox in bboxes_pdstrn:
x1 = int(bbox[0])
y1 = int(bbox[1])
x2 = int(bbox[2])
y2 = int(bbox[3])
trm_img = image[y1:y2, x1:x2]
pdstrn_trm_imges.append([trm_img])
except:
print("歩行者信号機のトリミングを試みましたが失敗しました")
trm_imges_dict["traffic_signal"] = traffic_trm_imges
trm_imges_dict["pedestrian_signal"] = pdstrn_trm_imges
if len(trm_imges_dict["traffic_signal"]) + len(
trm_imges_dict["pedestrian_signal"]) > 0:
for obj_name in ["traffic_signal", "pedestrian_signal"]:
mass_list = []
bboxes_info = bboxes_dict[obj_name]
res_data = extract_color_info(trm_imges_dict[obj_name])
#print("(r, g, b, h, s, v): ", res_data[0][4]) #Debug用
for input_data, bbox_info in zip(res_data, bboxes_info):
chunk_list = []
input_data = np.array(input_data[4])
pred, label_name = inference(input_data, clf)
#bboxes_info
chunk_list = bbox_info.tolist()
chunk_list.append(label_name)
chunk_list.append("信号色の確率値入れる")
mass_list.append(chunk_list)
#print(np.array(mass_list)) #Debug用
result[obj_name] = mass_list
image = draw_bboxes(image, result)
print(result)
del result
if arg == "video":
writer.write(image) # 画像を1フレーム分として書き込み
# 加工なし画像を表示する
cv2.imshow('Raw Frame', image)
# キー入力でqを押したら終了する
k = cv2.waitKey(1)
if k == ord('q'):
cv2.destroyAllWindows()
sys.exit()
else:
for img_path in imgs_path:
img_name = os.path.basename(img_path)
img = cv2.imread(img_path)
image, bboxes, _, _ = obj_inference(detector,
img,
count,
image_name=img_name,
flag=save_flag)
count += 1
break
# キャプチャをリリースして、ウィンドウをすべて閉じる
cap.release()
writer.release()
cv2.destroyAllWindows()
num = num + 1
area = (bbx[2] - bbx[0]) * (bbx[3] - bbx[1])
scales.append(float((bbx[2] - bbx[0]) / (bbx[3] - bbx[1])))
areas.append(area)
if num == 0:
newbboxes[keys] = np.zeros((0, 5), dtype=np.float32)
else:
areas = np.array(areas, dtype=np.float32)
areas = sorted(areas)
if len(areas) >= 5:
areas = np.delete(areas, 0, axis=0)
areas = np.delete(areas, 0, axis=0)
areas = np.delete(areas, len(areas) - 1, axis=0)
areas = np.delete(areas, len(areas) - 1, axis=0)
area_avg = areas.sum() / len(areas)
else:
area_avg = np.array(areas).sum() / len(areas)
scales = np.array(scales, dtype=np.float32)
scale_avg = scales.sum() / len(scales)
dets = []
for bbx in bboxes[keys]:
area = (bbx[2] - bbx[0]) * (bbx[3] - bbx[1])
scale = (bbx[2] - bbx[0]) / (bbx[3] - bbx[1])
if area <= 3.0 * area_avg and scale <= 1.5 * scale_avg:
dets.append(bbx)
dets = np.vstack(dets)
newbboxes[keys] = np.array(dets, dtype=np.float32)
image = draw_bboxes(image, newbboxes)
cv2.imwrite("demo_out.jpg", image)