Spaces:
Running
Running
File size: 12,161 Bytes
d2885a7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 | """
This file is the main runtime orchestrator for one person inside one frame.
It connects pose feature generation, posture classification, hand cropping, phone detection,
and final fusion into one clean modular flow. This is the clean replacement for the large
processOnePerson logic from your current project while keeping the same decision behavior.
"""
import time
from pathlib import Path
import cv2
import numpy as np
from src.components.distraction_fusion import DistractionFusion
from src.components.hand_cropper import HandCropper
from src.components.phone_detector import PhoneDetector
from src.components.pose_feature_generator import PoseFeatureGenerator
from src.components.posture_detector import PostureDetector
from src.config.constants import (
STATE_BACKSIDE,
STATE_NOT_USING,
STATE_OUT_OF_FRAME,
STATE_SUSPICIOUS,
STATE_TO_BE_CLASSIFIED,
)
from src.entity.config_entity import (
InferenceConfig,
MMPoseConfig,
PhoneDetectorConfig,
PostureModelConfig,
)
from src.utils.logger import get_logger
from src.utils.opencv_utils import (
crop_frame,
relative_to_absolute,
render_detection_rectangle,
resize_frame_to_square,
)
class RuntimeDetector:
"""
End-to-end runtime detector for one person.
"""
def __init__(
self,
mmpose_config: MMPoseConfig,
posture_model_config: PostureModelConfig,
phone_detector_config: PhoneDetectorConfig,
inference_config: InferenceConfig,
log_dir: Path | None = None,
log_level: str = "INFO",
) -> None:
self.mmpose_config = mmpose_config
self.posture_model_config = posture_model_config
self.phone_detector_config = phone_detector_config
self.inference_config = inference_config
self.logger = get_logger(
self.__class__.__name__, log_dir=log_dir, level=log_level
)
self.pose_feature_generator = PoseFeatureGenerator(
mmpose_config=mmpose_config,
posture_model_config=posture_model_config,
log_dir=log_dir,
log_level=log_level,
)
self.posture_detector = PostureDetector(
config=posture_model_config,
log_dir=log_dir,
log_level=log_level,
)
self.hand_cropper = HandCropper(
mmpose_config=mmpose_config,
phone_detector_config=phone_detector_config,
log_dir=log_dir,
log_level=log_level,
)
self.phone_detector = PhoneDetector(
config=phone_detector_config,
log_dir=log_dir,
log_level=log_level,
)
self.fusion = DistractionFusion(log_dir=log_dir, log_level=log_level)
self.posture_detector.load()
def _initial_state_checks(
self, keypoints: np.ndarray, xyxy: np.ndarray
) -> tuple[int, str]:
"""
Reproduce your current lightweight state machine checks
before posture classification.
"""
state = STATE_TO_BE_CLASSIFIED
score_text = ""
keypoint_score_threshold = (
self.inference_config.posture.keypoint_score_threshold
)
missing_threshold = (
self.inference_config.posture.out_of_frame_missing_keypoints_threshold
)
visible_keypoints = keypoints[:13, 2]
if np.sum(visible_keypoints < keypoint_score_threshold) >= missing_threshold:
return STATE_OUT_OF_FRAME, score_text
left_shoulder_x = keypoints[self.mmpose_config.keypoints.left_shoulder_index][0]
right_shoulder_x = keypoints[self.mmpose_config.keypoints.right_shoulder_index][
0
]
left_shoulder_score = keypoints[
self.mmpose_config.keypoints.left_shoulder_index
][2]
right_shoulder_score = keypoints[
self.mmpose_config.keypoints.right_shoulder_index
][2]
backside_ratio = (left_shoulder_x - right_shoulder_x) / (
(xyxy[2] - xyxy[0]) + np.finfo(np.float32).eps
)
if (
right_shoulder_score > keypoint_score_threshold
and left_shoulder_score > keypoint_score_threshold
and backside_ratio < self.inference_config.posture.backside_ratio_threshold
):
numeric_value = (
(right_shoulder_score + left_shoulder_score) / 2.0 + 1.0
) / 2.0
score_text = f"{numeric_value:.2f}"
return STATE_BACKSIDE, score_text
return state, score_text
def _run_posture_stage(self, keypoints: np.ndarray) -> tuple[int, str]:
"""
Convert keypoints to pose features and run posture classifier.
"""
feature_tensor = self.pose_feature_generator.build_feature_tensor(
keypoints, normalize=True
)
posture_result = self.posture_detector.predict(feature_tensor)
state = (
STATE_SUSPICIOUS if posture_result["class_signal"] == 1 else STATE_NOT_USING
)
score_text = posture_result["score_text"]
return state, score_text
def _run_phone_stage(
self,
frame: np.ndarray,
original_frame: np.ndarray,
keypoints: np.ndarray,
xyxy: np.ndarray,
) -> tuple[bool, np.ndarray | None]:
"""
Crop hands, run phone detection, and return phone status.
"""
use_trained_model = (
self.phone_detector_config.inference.use_trained_model_by_default
)
primary_crop, secondary_crop, spare_ratio = (
self.hand_cropper.get_priority_hand_crops(
frame=original_frame,
keypoints=keypoints,
xyxy=xyxy,
)
)
for crop_index, crop_item in enumerate([primary_crop, secondary_crop]):
if crop_item is None:
continue
hand_frame, hand_xyxy = crop_item
if hand_frame is None or hand_xyxy is None or hand_frame.size == 0:
continue
subframe_wh = (
abs(hand_xyxy[2] - hand_xyxy[0]),
abs(hand_xyxy[3] - hand_xyxy[1]),
)
resized_hand = resize_frame_to_square(
frame=hand_frame,
edge_length=(
self.phone_detector_config.inference.image_size
if use_trained_model
else 640
),
ratio_threshold=0.5625,
)
rgb_hand = cv2.cvtColor(resized_hand, cv2.COLOR_BGR2RGB)
phone_result = self.phone_detector.predict(
rgb_hand, use_trained=use_trained_model
)
render_detection_rectangle(
frame=frame,
text=f"Hand {crop_index}",
xyxy=hand_xyxy,
color="green" if not phone_result["detected"] else "pink",
)
if phone_result["detected"]:
relative_xyxy = np.array(
phone_result["relative_xyxy"], dtype=np.float32
)
from_mother_wh = (
(
self.phone_detector_config.inference.image_size
if use_trained_model
else 640
),
(
self.phone_detector_config.inference.image_size
if use_trained_model
else 640
),
)
absolute_xyxy = relative_to_absolute(
from_mother_wh=from_mother_wh,
to_mother_wh=subframe_wh,
from_child_xyxy=relative_xyxy,
to_mother_xy=(hand_xyxy[0], hand_xyxy[1]),
)
render_detection_rectangle(
frame=frame,
text=phone_result["text"],
xyxy=absolute_xyxy,
color="pink",
)
return True, None
if spare_ratio < self.inference_config.phone.spare_ratio_threshold:
break
return False, None
def process_one_person(
self,
frame: np.ndarray,
original_frame: np.ndarray,
keypoints: np.ndarray,
xyxy: np.ndarray,
runtime_parameters: dict,
) -> dict:
"""
Main entrypoint for one person in one frame.
Returns a structure similar to your current runtime response.
"""
start_posture = time.time()
initial_state, initial_score_text = self._initial_state_checks(keypoints, xyxy)
if initial_state in {STATE_OUT_OF_FRAME, STATE_BACKSIDE}:
posture_state = initial_state
posture_score_text = initial_score_text
else:
posture_state, posture_score_text = self._run_posture_stage(keypoints)
posture_time = time.time() - start_posture
phone_detected = False
announced_face_frame = None
face_xyxy = None
phone_time = 0.0
if posture_state == STATE_SUSPICIOUS:
start_phone = time.time()
phone_detected, announced_face_frame = self._run_phone_stage(
frame=frame,
original_frame=original_frame,
keypoints=keypoints,
xyxy=xyxy,
)
phone_time = time.time() - start_phone
fusion_result = self.fusion.fuse(
base_state=posture_state,
posture_score_text=posture_score_text,
phone_detected=phone_detected,
)
render_detection_rectangle(
frame=frame,
text=f"{fusion_result['display_text']} {fusion_result['score_text']}".strip(),
xyxy=xyxy,
color=fusion_result["display_color"],
)
# Optional face crop announce, similar to your old runtime logic
if fusion_result["final_label"] == "distracted":
face_center = keypoints[self.mmpose_config.keypoints.face_center_index][:2]
bbox_width = abs(xyxy[2] - xyxy[0])
left_ear_x = keypoints[self.mmpose_config.keypoints.left_ear_index][0]
right_ear_x = keypoints[self.mmpose_config.keypoints.right_ear_index][0]
face_len = max(
abs(int((right_ear_x - left_ear_x) * 1.1)), int(0.3 * bbox_width)
)
face_frame, face_xyxy = crop_frame(
original_frame, face_center, (face_len, face_len)
)
if face_frame is not None and face_xyxy is not None:
last_announce_time = runtime_parameters.get(
"time_last_announce_face", 0.0
)
current_frame_time = runtime_parameters.get(
"time_last_record_framerate", time.time()
)
if (
current_frame_time - last_announce_time
> self.inference_config.phone.face_announce_interval_seconds
):
announced_face_frame = face_frame
render_detection_rectangle(
frame=frame,
text="Face",
xyxy=face_xyxy,
color="white",
)
result = {
"performance": (posture_time, phone_time),
"announced_face_frame": announced_face_frame,
"posture": fusion_result["final_label"],
"phone": phone_detected,
"state": fusion_result["state"],
"display_text": fusion_result["display_text"],
"score_text": fusion_result["score_text"],
# "face_xyxy": face_xyxy.tolist() if face_xyxy is not None else None,
"face_xyxy": (
face_xyxy.tolist()
if hasattr(face_xyxy, "tolist")
else face_xyxy if face_xyxy is not None else None
),
}
self.logger.info("Runtime one-person result: %s", result)
return result
|