def framewise_recognize(pose, pretrained_model):
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
joints_norm_per_frame = np.array(pose[-1])
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0 here).
detections = [Detection(bbox, 1.0, feature) for bbox, feature in zip(bboxes, features)]
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
tracker.predict()
tracker.update(detections)
trk_result = []
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
trk_result.append([bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
trk_id = 'ID-' + str(trk.track_id)
cv.putText(frame, trk_id, (int(bbox[0]), int(bbox[1]-45)), cv.FONT_HERSHEY_SIMPLEX, 0.8, trk_clr, 3)
for d in trk_result:
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
# id = int(d[4])
try:
tmp = np.array([abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
j = 0
if joints_norm_per_frame.size > 0:
joints_norm_single_person = joints_norm_per_frame[j*36:(j+1)*36]
joints_norm_single_person = np.array(joints_norm_single_person).reshape(-1, 36)
pred = np.argmax(pretrained_model.predict(joints_norm_single_person))
init_label = Actions(pred).name
cv.putText(frame, init_label, (xmin + 80, ymin - 45), cv.FONT_HERSHEY_SIMPLEX, 1, trk_clr, 3)
if init_label == 'fall_down':
cv.putText(frame, 'WARNING: someone is falling down!', (20, 60), cv.FONT_HERSHEY_SIMPLEX,
1.5, (0, 0, 255), 4)
cv.rectangle(frame, (xmin - 10, ymin - 30), (xmax + 10, ymax), trk_clr, 2)
return frame
def get_pose_dict(pose, cur_pose, frame_count, cur_box):
cur_box.clear()
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
joints_norm_per_frame = np.array(pose[-1]) #720p下的所有关节点坐标(相对坐标)
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
# 进行非极大抑制
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# 调用tracker并实时更新
tracker.predict()
tracker.update(detections)
# 记录track的结果，包括bounding boxes及其ID
trk_result = []
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
for d in trk_result:
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
try:
# xcenter是一帧图像中所有human的1号关节点（neck）的x坐标值
# 通过计算track_box与human的xcenter之间的距离，进行ID的匹配
tmp = np.array([abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
# 若当前帧无human，默认j=0（无效）
j = 0
if joints_norm_per_frame.size > 0 and frame_count % 5 == 0: ####################################
tmp = [frame_count]
tmp.extend(joints_norm_per_frame[j * 36:(j + 1) * 36])
cur_pose[d[4]] = tmp
cur_box[d[4]] = [xmin, ymin, xmax, ymax]
elif joints_norm_per_frame.size > 0:
cur_box[d[4]] = [xmin, ymin, xmax, ymax]
return 1
def framewise_recognize(pose, pretrained_model):
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
joints_norm_per_frame = np.array(pose[-1])
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
tracker.predict()
tracker.update(detections)
trk_result = []
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
trk_id = 'ID-' + str(trk.track_id)
cv.putText(frame, trk_id, (int(bbox[0]), int(bbox[1] - 45)),
cv.FONT_HERSHEY_SIMPLEX, 0.8, trk_clr, 3)
for d in trk_result:
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
# id = int(d[4])
try:
tmp = np.array([abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
j = 0
cv.rectangle(frame, (xmin - 10, ymin - 30), (xmax + 10, ymax),
trk_clr, 2)
return frame
def framewise_recognize(pose, pretrained_model):
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
joints_norm_per_frame = np.array(pose[-1])
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
# 비 최대 억제
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# 이전 tracker 실시간 추적 및 업데이트
tracker.predict()
tracker.update(detections)
# bounding boxes 및 해당 ID를 포함하는 track의 결과를 기록
trk_result = []
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
# track_ID 표시
trk_id = 'ID-' + str(trk.track_id)
cv.putText(frame, trk_id, (int(bbox[0]), int(bbox[1] - 45)),
cv.FONT_HERSHEY_SIMPLEX, 0.8, trk_clr, 3)
for d in trk_result:
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
try:
# xcenter는 이미지 프레임에서 모든 인간의 No.1 조인트 포인트임(neck）X좌표 값
# track_box와 human xcenter 사이의 거리를 계산하여 ID 매칭을 수행함
tmp = np.array([abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
# 현재 프래임에 사람이 없으면 기본값 j=0(무효)
j = 0
# 행동 분류 수행
if joints_norm_per_frame.size > 0:
joints_norm_single_person = joints_norm_per_frame[j *
36:(j + 1) *
36]
joints_norm_single_person = np.array(
joints_norm_single_person).reshape(-1, 36)
pred = np.argmax(
pretrained_model.predict(joints_norm_single_person))
init_label = Actions(pred).name
# Action 카테고리 표시
cv.putText(frame, init_label, (xmin + 80, ymin - 45),
cv.FONT_HERSHEY_SIMPLEX, 1, trk_clr, 3)
# 비정상적인 경고(under scene)
if init_label == 'fall_down':
cv.putText(frame, 'WARNING: someone is falling down!',
(20, 60), cv.FONT_HERSHEY_SIMPLEX, 1.5,
(0, 0, 255), 4)
# track_box 그리기
cv.rectangle(frame, (xmin - 10, ymin - 30), (xmax + 10, ymax),
trk_clr, 2)
return frame
def framewise_recognize(pose, pretrained_model):
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
joints_norm_per_frame = np.array(pose[-1])
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
# 进行非极大抑制
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# 调用tracker并实时更新
tracker.predict()
tracker.update(detections)
# 记录track的结果，包括bounding boxes及其ID
trk_result = []
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
# 标注track_ID
trk_id = 'ID-' + str(trk.track_id)
cv.putText(frame, trk_id, (int(bbox[0]), int(bbox[1] - 45)),
cv.FONT_HERSHEY_SIMPLEX, 0.8, trk_clr, 3)
for d in trk_result:
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
id = int(d[4]) # <<<<
try:
# xcenter is the x coordinate value of the joint point (neck) of all humans in a frame of image
# By calculating the distance between track_box and human xcenter, ID matching
tmp = np.array([abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
# no humans !!
j = 0
# faire la classification =========================================================
if joints_norm_per_frame.size > 0:
joints_norm_single_person = joints_norm_per_frame[j *
36:(j + 1) *
36]
#joints_norm_single_person = np.array(joints_norm_single_person).reshape(-1, 36)
#print("==========================================")
#print(id)
#print(len(joints_norm_single_person[0]))
#print(joints_norm_single_person[0])
#print("==========================================")
exists, ret = lstm_pred.update_ctx(id,
joints_norm_single_person)
# print(exists)
# print("----")
if exists:
print(f"predicting for {id}")
#print(np.array(lstm_pred.ctx[id]).reshape(-1, 4, 36).shape)
# الحكمة ممبعد تجي هنا
ret = np.array(ret).reshape(-1, 4, 36)
#print(pretrained_model.predict(ret))
pred = np.argmax(pretrained_model.predict(ret))
print(pred)
init_label = Actions(pred).name
# 显示动作类别
cv.putText(frame, init_label, (xmin + 80, ymin - 45),
cv.FONT_HERSHEY_SIMPLEX, 1, trk_clr, 3)
# 异常预警(under scene)
if init_label == 'fall_down':
cv.putText(frame, 'WARNING: someone is falling down!',
(20, 60), cv.FONT_HERSHEY_SIMPLEX, 1.5,
(0, 0, 255), 4)
# 画track_box
cv.rectangle(frame, (xmin - 10, ymin - 30), (xmax + 10, ymax),
trk_clr, 2)
return frame
def framewise_recognize(pose, pretrained_model, camera):
global last_init_label, capture_picture_count, start_time, fir, fir_ID, getSignal
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
joints_norm_per_frame = np.array(pose[-1])
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
# 进行非极大抑制，选择置信得分最高的边框并保留
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
# 调用tracker并实时更新
tracker.predict()
tracker.update(detections)
# 记录track的结果，包括bounding boxes及其ID
trk_result = []
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
# 标注track_ID
trk_id = 'ID-' + str(trk.track_id)
cv.putText(frame, trk_id, (int(bbox[0]), int(bbox[1] - 45)),
cv.FONT_HERSHEY_SIMPLEX, 0.8, trk_clr, 3)
# 找出第一个进入镜头的人的ID
if fir == 'true':
fir = 'false'
fir_ID = trk.track_id
print("First person ID:{}".format(fir_ID))
for d in trk_result:
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
id = d[4]
try:
# xcenter是一帧图像中所有human的1号关节点（neck）的x坐标值
# 通过计算track_box与human的xcenter之间的距离，进行ID的匹配
tmp = np.array([abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
# 若当前帧无human，默认j=0（无效）
j = 0
# 进行动作分类
if joints_norm_per_frame.size > 0:
joints_norm_single_person = joints_norm_per_frame[j *
36:(j + 1) *
36]
joints_norm_single_person = np.array(
joints_norm_single_person).reshape(-1, 36)
pred = np.argmax(
pretrained_model.predict(joints_norm_single_person))
init_label = Actions(pred).name
# 显示动作类别
cv.putText(frame, init_label, (xmin + 80, ymin - 45),
cv.FONT_HERSHEY_SIMPLEX, 1, trk_clr, 3)
# 检测到指定动作，拍摄照片,瞬间拍照
# if (init_label == 'wave' and last_init_label !='wave' and capture_picture_count <= 10 and time.time() - start_time > 2 and id == fir_ID):
# ret, frame = camera.read()
# filename = "img/getPicutre-{0}.jpg".format(capture_picture_count)
# cv.imwrite(filename, frame)
# print("Image saved at {}".format(filename))
# capture_picture_count=capture_picture_count+1
# start_time=time.time()
# 检测到指定动作，在1.5s后拍照
if (init_label == 'wave' and getSignal == 'false'):
getSignal = 'true'
print("Prepare to get image\n")
start_time = time.time()
if (getSignal == 'true' and time.time() - start_time > 1.5):
getSignal = 'false'
ret, frame = camera.read()
filename = "img/getPicutre-{0}.jpg".format(
capture_picture_count)
cv.imwrite(filename, frame)
print("Image saved at {}\n".format(filename))
capture_picture_count = capture_picture_count + 1
# 检测到指定动作，改变相机参数，待实现
# else if(init_label == 'XXX' and last_init_label !='XXX' and time.time() - start_time > 1 and id == fir_ID)
# myGain=cap.get(cv2.CAP_PROP_GAIN)
# if(myGain < 20)
# cap.set(cv2.CAP_PROP_GAIN, myGain+2) #调整增益
last_init_label = init_label
# 画track_box
cv.rectangle(frame, (xmin - 10, ymin - 30), (xmax + 10, ymax),
trk_clr, 2)
return frame
def show_camera(self):
start = time.time()
ret, frame = self.cap.read()
# image_h, image_w = frame.shape[:2] # (720x960)
#show = cv2.resize(frame, (settings.winWidth, settings.winHeight))
show = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
if ret:
if self.__flag_mode == 1:
self.infoBox.setText(u'Pose estimation')
humans = my_detector.detect(show)
#skelsInfo = SkeletonDetector.humans_to_skelsInfo(humans)
#print('SkelsInfo: ', skelsInfo)
my_detector.draw(show, humans)
# for ith_skel in range(0, len(skelsInfo)):
# skeleton = SkeletonDetector.get_ith_skeleton(skelsInfo, ith_skel)
# if ith_skel:#ith_skel == 0:
elif self.__flag_mode == 2:
self.infoBox.setText(u'Multiplayer tracking')
prediced_label = ''
humans = my_detector.detect(show)
my_detector.draw(show, humans)
skelsInfo, skelsInfo_choose, bboxes = SkeletonDetector.humans_to_skelsInfo(
humans, joint_choose)
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0 here
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
# Non-maximal suppression 进行非极大抑制
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(
boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# Call tracker and update in real time 调用tracker并实时更新
tracker.predict()
tracker.update(detections)
# Record track results, including bounding boxes and their ID 记录track的结果，包括bounding boxes及其ID
trk_result = []
#trk.track_id = 1
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
print('bbox: ', bbox)
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
# Đánh dấu track_ID
trk_id = 'ID-' + str(trk.track_id)
print('track ID: ', trk_id)
cv2.putText(show, trk_id,
(int(bbox[0]), int(bbox[1] - 45)),
cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 0, 0),
3) ################################
for d in trk_result: # gom nhieu bboxes
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
label = int(d[4])
cv2.rectangle(show, (xmin, ymin), (xmax, ymax),
(int(settings.c[label % 32, 0]),
int(settings.c[label % 32, 1]),
int(settings.c[label % 32, 2])), 4)
elif self.__flag_mode == 3:
self.infoBox.setText(u'Sign language recognition')
humans = my_detector.detect(show)
my_detector.draw(show, humans)
#skelsInfo_choose = SkeletonDetector.humans_to_skelsInfo_choose(humans, joint_choose) # cac skeleton_choose
skelsInfo, skelsInfo_choose, bboxes = SkeletonDetector.humans_to_skelsInfo(
humans, joint_choose) # cac skeleton_full
# if len(skelsInfo) == 0:
# self.text = ''
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0 here
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
# Non-maximal suppression
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(
boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
# Call tracker and update in real time
tracker.predict()
tracker.update(detections)
# Record track results, including bounding boxes and their ID
trk_result = []
#trk.track_id = 1
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
#print('bbox: ', bbox)
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
# Đánh dấu track_ID
trk_id = 'ID-' + str(trk.track_id)
#print('track ID: ', trk_id)
cv2.putText(show, trk_id,
(int(bbox[0]), int(bbox[1] - 45)),
cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 0, 0),
3) ################################
list_label = []
for d in trk_result: # gom nhieu bboxes
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
label = int(d[4])
list_label.append(label)
cv2.rectangle(show, (xmin, ymin), (xmax, ymax),
(int(settings.c[label % 32, 0]),
int(settings.c[label % 32, 1]),
int(settings.c[label % 32, 2])), 4)
try:
# xcenter là giá trị tọa độ x của tất cả các khớp (cổ) của con người trong một khung hình ảnh
# Khớp ID bằng cách tính khoảng cách giữa track_box và xcenter của con người
tmp = np.array(
[abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
# Nếu không có người trong khung hiện tại, mặc định j = 0 (không hợp lệ)
j = 0
if skelsInfo_choose:
skeleton_choose = skelsInfo_choose[j]
skeleton = skelsInfo[j]
skeleton = np.array(skeleton[1:1 + 18 * 2])
#skeleton =
if len(skeleton_choose) == 20:
#skeleton_choose = np.array(skelsInfo_choose).reshape(-1,20)
prediced_label, acc, vni_prediced_label = classifier.predict(
skeleton_choose)
if acc > 0.94:
if label not in list(self.dict):
self.dict[label] = vni_prediced_label
self.dict_check[
label] = vni_prediced_label
if label in list(self.dict) and (
vni_prediced_label !=
self.dict_check[label]):
self.dict[label] = self.dict[
label] + ' ' + vni_prediced_label
self.dict_check[
label] = vni_prediced_label
#print('prediced_label: ', prediced_label)
cv2.putText(show, prediced_label,
(xmin, ymin - 10),
cv2.FONT_HERSHEY_SIMPLEX, 1.5,
(255, 0, 0), 4, cv2.LINE_AA)
if len(self.dict_check) > 0:
for i in list(self.dict_check):
if i not in list_label:
del self.dict_check[i]
del self.dict[i]
self.text = ''
for i in list(self.dict):
self.text = self.text + 'ID-' + str(
i) + ': ' + self.dict[i] + '\n'
self.textBox.setText(self.text)
end = time.time()
self.fps = 1. / (end - start)
cv2.putText(show, 'FPS: %.2f' % self.fps, (30, 30),
cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
showImage = QtGui.QImage(show, show.shape[1], show.shape[0],
QtGui.QImage.Format_RGB888)
self.label_show_camera.setPixmap(
QtGui.QPixmap.fromImage(showImage))
def framewise_recognize(pose, pretrained_model):
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
joints_norm_per_frame = np.array(pose[-1]) #720p下的所有关节点坐标(相对坐标)
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame, bboxes)
# score to 1.0 here).
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
# 进行非极大抑制
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices]
#先 detection再track
# 调用tracker并实时更新
tracker.predict()
tracker.update(detections)
# 记录track的结果，包括bounding boxes及其ID
trk_result = []
for trk in tracker.tracks:
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh()
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id])
#bbox[0]:ROI左上角x bbox[1]:ROI左上角y,bbox[2]:ROI宽度，bbox[3]:ROI高度
# 标注track_ID
trk_id = 'ID-' + str(trk.track_id)
cv.putText(frame, trk_id, (int(bbox[0]), int(bbox[1] - 45)),
cv.FONT_HERSHEY_SIMPLEX, 0.8, trk_clr, 3)
for d in trk_result:
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
# id = int(d[4])
try:
# xcenter是一帧图像中所有human的1号关节点（neck）的x坐标值
# 通过计算track_box与human的xcenter之间的距离，进行ID的匹配
tmp = np.array([abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
# 若当前帧无human，默认j=0（无效）
j = 0
# 进行动作分类 previously tried
if joints_norm_per_frame.size > 0:
joints_norm_single_person = joints_norm_per_frame[j *
36:(j + 1) *
36]
joints_norm_single_person = np.array(
joints_norm_single_person).reshape(-1, 36)
#joints_norm_single_person = np.array(joints_norm_single_person).reshape(-1, 6,6)##changed be caution
pred = np.argmax(
pretrained_model.predict(joints_norm_single_person))
init_label = Actions(pred).name
# 显示动作类别
cv.putText(frame, init_label, (xmin + 80, ymin - 45),
cv.FONT_HERSHEY_SIMPLEX, 1, trk_clr, 3)
# 异常预警(under scene)
if init_label == 'fall_down':
cv.putText(frame, 'WARNING: someone is falling down!',
(20, 60), cv.FONT_HERSHEY_SIMPLEX, 1.5,
(0, 0, 255), 4)
# 画track_box
cv.rectangle(frame, (xmin - 10, ymin - 30), (xmax + 10, ymax),
trk_clr, 2)
return frame
def framewise_recognize(pose, pretrained_model):
global fall_num
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
#frame是已经标记好骨骼点的图片
joints_norm_per_frame = np.array(pose[-1])
# Ano_list = draw_region_detection(frame) #得到检测范围坐标
# boxss = bboxes #box的格式为tlxmin tlymin width height
# boxs = []
# # # 标定检测范围的代码
# for i in range(len(boxss)):
# bb = boxss[i]
# if bb[0] > Ano_list[0] and bb[0] < Ano_list[2]: # 判断要检测的范围 288,189
# boxs.append(boxss[i]) # 也就是不检测 不匹配
# bboxes = boxs
#追踪部分
if bboxes:
bboxes = np.array(bboxes)
features = encoder(frame,
bboxes) #获得128维的特征的编码 deep_sort中的 #每个box相当于一个patch
#将patch送到deep sort的网络中去进行编码
# score to 1.0 here).
# 特征和bboxes对应 这个句话也就是转换了一下数据类型 也相当于建立了这个类型的实例
detections = [
Detection(bbox, 1.0, feature)
for bbox, feature in zip(bboxes, features)
]
# 进行非极大抑制
boxes = np.array([d.tlwh for d in detections])
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap,
scores)
detections = [detections[i] for i in indices] #选取最终的都的目标坐标
# 调用tracker并实时更新
# 第一帧作为初始化 后面根据前帧 后帧的距离判断 达到跟踪的目的
tracker.predict() #得到已有目标预测后的mean 和 covariance 用来作为判断依据
tracker.update(detections) #是不是更新了跟踪的目标集合 检测到的目标集合
# 记录track的结果，包括bounding boxes及其ID
trk_result = []
#好多利用for直接实例化的例子用法啊
for trk in tracker.tracks: # 这是一个list 啊 对每个逐一判断
if not trk.is_confirmed() or trk.time_since_update > 1:
continue
bbox = trk.to_tlwh() #tlah - > to top left coord and weight,height
trk_result.append(
[bbox[0], bbox[1], bbox[2], bbox[3], trk.track_id]) #t,l
# 标注track_ID
trk_id = 'ID-' + str(trk.track_id)
# cv.putText(frame, trk_id, (int(bbox[0]), int(bbox[1]-45)), cv.FONT_HERSHEY_SIMPLEX, 0.8, trk_clr, 2)
for d in trk_result:
xmin = int(d[0])
ymin = int(d[1])
xmax = int(d[2]) + xmin
ymax = int(d[3]) + ymin
# id = int(d[4])
try:
# xcenter是一帧图像中所有human的1号关节点（neck）的x坐标值
# 通过计算track_box与human的xcenter之间的距离，进行ID的匹配
tmp = np.array([abs(i - (xmax + xmin) / 2.) for i in xcenter])
j = np.argmin(tmp)
except:
# 若当前帧无human，默认j=0（无效）
j = 0
# 进行动作分类
if joints_norm_per_frame.size > 0: #每个人具有36个维度 18个位置的关节点
joints_norm_single_person = joints_norm_per_frame[j *
36:(j + 1) *
36]
joints_norm_single_person = np.array(
joints_norm_single_person).reshape(-1, 36)
pred = np.argmax(
pretrained_model.predict(
joints_norm_single_person)) #只是keras里的predict
init_label = Actions(pred).name #得到enum标签的名称
# 显示动作类别 ymin - 40
# cv.putText(frame, init_label, (xmin + 30, ymax + 20), cv.FONT_HERSHEY_SIMPLEX, 1, trk_clr, 2)
# 异常预警(under scene)
if init_label == 'fall_down':
fall_num += 1
if fall_num > 2: #至少连续2帧出现异常时，才能出现报错的情况
fall_num = 2
if fall_num == 2:
cv.putText(frame, 'WARNING: someone is Falling down!',
(20, 60), cv.FONT_HERSHEY_SIMPLEX, 1,
(0, 0, 255), 3)
else:
fall_num = 0
print('fall_num:{}'.format(fall_num))
# 画track_box
# cv.rectangle(frame, (xmin - 10, ymin - 30), (xmax + 10, ymax), trk_clr, 2)
return frame