def _main(cap_src):
cap = cv2.VideoCapture(cap_src)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
face_d = FaceDetector()
sess = onnxruntime.InferenceSession(
f'{root_path}/pretrained/fsanet-1x1-iter-688590.onnx')
sess2 = onnxruntime.InferenceSession(
f'{root_path}/pretrained/fsanet-var-iter-688590.onnx')
print('Processing frames, press q to exit application...')
while True:
ret, frame = cap.read()
if (not ret):
print(
'Could not capture a valid frame from video source, check your cam/video value...'
)
break
#get face bounding boxes from frame
face_bb = face_d.get(frame)
for (x1, y1, x2, y2) in face_bb:
face_roi = frame[y1:y2 + 1, x1:x2 + 1]
#preprocess headpose model input
face_roi = cv2.resize(face_roi, (64, 64))
face_roi = face_roi.transpose((2, 0, 1))
face_roi = np.expand_dims(face_roi, axis=0)
face_roi = (face_roi - 127.5) / 128
face_roi = face_roi.astype(np.float32)
#get headpose
res1 = sess.run(["output"], {"input": face_roi})[0]
res2 = sess2.run(["output"], {"input": face_roi})[0]
yaw, pitch, roll = np.mean(np.vstack((res1, res2)), axis=0)
draw_axis(frame,
yaw,
pitch,
roll,
tdx=(x2 - x1) // 2 + x1,
tdy=(y2 - y1) // 2 + y1,
size=50)
#draw face bb
# cv2.rectangle(frame,(x1,y1),(x2,y2),(0,255,0),2)
cv2.imshow('Frame', frame)
key = cv2.waitKey(1) & 0xFF
if (key == ord('q')):
break
def _main(cap_src):
prev_data = []
data = []
mode2 = NodShakeMode(prev_data, data)
hmm_model = NodShakeHMM(maxlen=12)
cap = cv2.VideoCapture(cap_src)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
face_d = FaceDetector()
eye_model = f'{root_path}/pretrained/haarcascades/haarcascade_eye_tree_eyeglasses.xml'
sess = onnxruntime.InferenceSession(f'{root_path}/pretrained/fsanet-1x1-iter-688590.onnx')
sess2 = onnxruntime.InferenceSession(f'{root_path}/pretrained/fsanet-var-iter-688590.onnx')
tracking_dict = {k:deque(maxlen=MAXLEN) for k in ['yaw', 'pitch', 'roll']}
font = cv2.FONT_HERSHEY_SIMPLEX
collect_head_pose = None
prev_head_pose_hmm = 'stationary'
print('Processing frames, press q to exit application...')
while True:
ret,frame = cap.read()
if(not ret):
print('Could not capture a valid frame from video source, check your cam/video value...')
break
#get face bounding boxes from frame
if COLLECT_DATA:
face_bb = face_d.get(frame)
head_pose = ""
else:
face_bb, data = face_d.detect_face_and_eyes_enhanced(frame, cv2.CascadeClassifier(eye_model))
# mode2.set_data(prev_data, data)
# head_pose = mode2.apply()
# prev_data.append(data)
for (x1, y1, x2, y2) in face_bb:
if x1 < 0 or y1 < 0 or x2 < 0 or y2 < 0:
continue
face_roi = frame[y1:y2+1,x1:x2+1]
#preprocess headpose model input
face_roi = cv2.resize(face_roi,(64,64))
face_roi = face_roi.transpose((2,0,1))
face_roi = np.expand_dims(face_roi,axis=0)
face_roi = (face_roi-127.5)/128
face_roi = face_roi.astype(np.float32) # -> (1, 3, 64, 64)
#get headpose
res1 = sess.run(["output"], {"input": face_roi})[0]
res2 = sess2.run(["output"], {"input": face_roi})[0]
yaw, pitch, roll = np.mean(np.vstack((res1, res2)), axis=0)
tracking_dict["yaw"].append(yaw)
tracking_dict["pitch"].append(pitch)
tracking_dict["roll"].append(roll)
if not COLLECT_DATA:
data.add_euler_angles(yaw, pitch, roll)
mode2.set_data(prev_data, data)
head_pose = mode2.apply()
if data.x1:
prev_data.append(data)
hmm_model.add_data(data)
new_head_pose_hmm = hmm_model.determine_pose()
if new_head_pose_hmm == 'stationary' or prev_head_pose_hmm == new_head_pose_hmm:
head_pose_hmm = new_head_pose_hmm
else:
head_pose_hmm = prev_head_pose_hmm
prev_head_pose_hmm = new_head_pose_hmm
# head_pose = ''
print(datetime.datetime.now(), yaw, pitch, roll)
# print(np.std(tracking_dict["yaw"]), np.std(tracking_dict["pitch"]), np.std(tracking_dict["roll"]))
# # Nodding is when pitch is changing fairly sinusoidal while roll and yaw stays relatively consistent
# if np.std(tracking_dict["yaw"]) < 3 and np.std(tracking_dict["roll"]) < 3 and np.std(tracking_dict["pitch"]) > 3:
# head_pose = 'NOD'
draw_axis(frame, yaw, pitch, roll, tdx=(x2-x1)//2+x1, tdy=(y2-y1)//2+y1, size=50)
# cv2.putText(frame, head_pose, (x1, y1), font, 2, (0, 0, 0), 3, cv2.LINE_AA)
cv2.putText(frame, head_pose_hmm, (x2, y2), font, 2, (0, 0, 0), 3, cv2.LINE_AA)
#draw face bb
cv2.rectangle(frame,(x1,y1),(x2,y2),(0,255,0),2)
if COLLECT_DATA and len(tracking_dict["yaw"]) == MAXLEN:
if not collect_head_pose:
collect_head_pose = random.choice(HEAD_POSE)
else:
df = pd.DataFrame.from_dict(tracking_dict)
df.to_csv(
'collected_data/%s_%s_%s'%(
collect_head_pose,
MAXLEN,
start_idx[collect_head_pose]
),
index=False,
)
start_idx[collect_head_pose] += 1
collect_head_pose = random.choice(HEAD_POSE)
for angle in EULER_ANGLES:
tracking_dict[angle].clear()
input("Enter to continue and do head pose: %s"%collect_head_pose)
frame = ResizeWithAspectRatio(frame, width=1000)
cv2.imshow('Frame', frame)
key = cv2.waitKey(1)&0xFF
if(key == ord('q')):
break