def plot_results(detector, pose, img, category, logging=True):
pose_img = cv2.cvtColor(img, cv2.COLOR_BGRA2BGR)
h, w = img.shape[0], img.shape[1]
count = detector.get_object_count()
if logging:
logger.info(f'object_count={count}')
for idx in range(count):
obj = detector.get_object(idx)
# print result
if logging:
logger.info(f'+ idx={idx}')
logger.info(
f' category={obj.category}[ {category[obj.category]} ]')
logger.info(f' prob={obj.prob}')
logger.info(f' x={obj.x}')
logger.info(f' y={obj.y}')
logger.info(f' w={obj.w}')
logger.info(f' h={obj.h}')
top_left = (int(w * obj.x), int(h * obj.y))
bottom_right = (int(w * (obj.x + obj.w)), int(h * (obj.y + obj.h)))
text_position = (int(w * obj.x) + 4, int(h * (obj.y + obj.h) - 8))
# update image
color = hsv_to_rgb(256 * obj.category / len(category), 255, 255)
fontScale = w / 512.0
cv2.rectangle(img, top_left, bottom_right, color, 4)
cv2.putText(img, category[obj.category], text_position,
cv2.FONT_HERSHEY_SIMPLEX, fontScale, color, 1)
CATEGORY_PERSON = 0
if obj.category != CATEGORY_PERSON:
continue
# pose detection
px1, py1, px2, py2 = keep_aspect(top_left, bottom_right, pose_img,
pose)
crop_img = pose_img[py1:py2, px1:px2, :]
offset_x = px1 / img.shape[1]
offset_y = py1 / img.shape[0]
scale_x = crop_img.shape[1] / img.shape[1]
scale_y = crop_img.shape[0] / img.shape[0]
detections = compute(pose, crop_img, offset_x, offset_y, scale_x,
scale_y)
cv2.rectangle(img, (px1, py1), (px2, py2), color, 1)
display_result(img, detections)
return img
def pose_estimation(detector, pose, img):
pose_img = cv2.cvtColor(img, cv2.COLOR_BGRA2BGR)
h, w = img.shape[0], img.shape[1]
count = detector.get_object_count()
pose_detections = []
for idx in range(count):
obj = detector.get_object(idx)
top_left = (int(w * obj.x), int(h * obj.y))
bottom_right = (int(w * (obj.x + obj.w)), int(h * (obj.y + obj.h)))
CATEGORY_PERSON = 0
if obj.category != CATEGORY_PERSON:
pose_detections.append(None)
continue
px1, py1, px2, py2 = keep_aspect(top_left, bottom_right, pose_img,
pose)
crop_img = pose_img[py1:py2, px1:px2, :]
offset_x = px1 / img.shape[1]
offset_y = py1 / img.shape[0]
scale_x = crop_img.shape[1] / img.shape[1]
scale_y = crop_img.shape[0] / img.shape[0]
detections = compute(pose, crop_img, offset_x, offset_y, scale_x,
scale_y)
pose_detections.append(detections)
return pose_detections
def action_recognition(box, input_image, pose, detector, model, data):
if args.arch == "lw_human_pose":
bbox_xywh, cls_conf, cls_ids = get_detector_result_lw_human_pose(
pose, input_image.shape[0], input_image.shape[1], get_all=True)
idx = -1
min_d = 32768
for i in range(pose.get_object_count()):
target = xywh_to_xyxy(bbox_xywh[i], input_image.shape[0],
input_image.shape[1])
d = math.sqrt((target[0] - box[0])**2 + (target[1] - box[1])**2)
if d < min_d:
min_d = d
idx = i
if idx == -1:
return "-", None
person = pose.get_object_pose(idx)
else:
bbox_xywh, cls_conf, cls_ids = get_detector_result(
detector, input_image.shape[0], input_image.shape[1])
px1, py1, px2, py2 = keep_aspect((box[0], box[1]), (box[2], box[3]),
input_image, pose)
crop_img = input_image[py1:py2, px1:px2, :]
offset_x = px1 / input_image.shape[1]
offset_y = py1 / input_image.shape[0]
scale_x = crop_img.shape[1] / input_image.shape[1]
scale_y = crop_img.shape[0] / input_image.shape[0]
detections = compute(pose, crop_img, offset_x, offset_y, scale_x,
scale_y)
person = detections
openpose_keypoints = ailia_to_openpose(person)
frame = np.expand_dims(openpose_keypoints, axis=1)
frame = pose_postprocess(frame)
for i in range(TIME_RANGE - 1):
data[:,
i, :] = data[:, i +
1, :] #data: (ailia.POSE_KEYPOINT_CNT,TIME_RANGE,3)
data[:, TIME_RANGE - 1, :] = frame[:, 0, :]
zero_cnt = 0
for i in range(TIME_RANGE):
if np.sum(data[:, i, :]) == 0:
zero_cnt = zero_cnt + 1
if zero_cnt >= 1:
return "-", person
data_rgb = data.transpose((2, 1, 0))
data_rgb = data_rgb[:2,
...] # May need to be removed if input for action model changed
data_rgb = np.expand_dims(data_rgb, axis=3)
data_rgb.shape = (1, ) + data_rgb.shape
model.set_input_shape(data_rgb.shape)
action = model.predict(data_rgb)
action = softmax(action)
max_prob = 0
class_idx = 0
for i in range(len(LABELS)):
if max_prob <= action[0][i]:
max_prob = action[0][i]
class_idx = i
return LABELS[class_idx] + " " + str(int(max_prob * 100) / 100), person