[2D Object Detection] Carla Simulator에서의 2D dataset 추출.

[김찬우]

데이터셋 추출 환경

Carla Simulator 0.9.15

Anaconda 가상 환경

Unreal Engine 4

1. CarlaUE4.exe 실행하여 서버 유지

2. manual_control.py 실행하여 ego 차량 소환 (데이터셋 추출 시에는 'p'로  autopilot 모드 실행.)

manual_control 조작법

날씨 변경 : Ctrl + C
차량 변경 : backspace
시점 변경 : Tab
오토파일럿 : P
카메라 변경 : N
조종 : W A S D
기어 변경 : M 키로 직접 변경 변환 후 , / . 로 기어 down, up 조종

TrafficObject 

데이터 수집

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

import math import random import time import queue import numpy as np import cv2 import json import os import logging from datetime import datetime import glob import sys   # CARLA Python API 경로 설정 try:     sys.path.append(glob.glob('../carla/dist/carla-*%d.%d-%s.egg' % (         sys.version_info.major,         sys.version_info.minor,         'win-amd64' if os.name == 'nt' else 'linux-x86_64'))[0]) except IndexError:     pass   import carla   # 로깅 설정 logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')   # === 설정 === IM_WIDTH = 800 IM_HEIGHT = 600 SAVE_INTERVAL = 1.0  # 초 단위 VISIBLE_THRESHOLD = 0.05  # 5% 이상 보일 때만 저장 CLASS_IDS = {     'vehicle': {'carla_ids': [10, 14, 15, 18], 'yolo_id': 0},  # 차량, 트럭, 오토바이 등     'pedestrian': {'carla_ids': [12], 'yolo_id': 1}  # 보행자 } OUTPUT_DIR = "yolo_dataset" COCO_DIR = "coco_dataset" os.makedirs(os.path.join(OUTPUT_DIR, "images"), exist_ok=True) os.makedirs(os.path.join(OUTPUT_DIR, "labels"), exist_ok=True) os.makedirs(os.path.join(COCO_DIR, "images"), exist_ok=True)   # === COCO JSON 초기화 === coco_json = {     "info": {         "description": "CARLA Vehicle and Pedestrian Detection Dataset",         "version": "1.0",         "year": 2025,         "date_created": datetime.now().strftime("%Y-%m-%d")     },     "images": [],     "annotations": [],     "categories": [         {"id": 0, "name": "vehicle", "supercategory": "vehicle"},         {"id": 1, "name": "pedestrian", "supercategory": "pedestrian"}     ] } annotation_id = 1 image_id = 1   # === client 및 world 초기화 === client = carla.Client("localhost", 2000) client.set_timeout(10.0) world = client.get_world() bp_lib = world.get_blueprint_library()   # === ego vehicle 찾기 === vehicle = None for actor in world.get_actors().filter('vehicle.*'):     if actor.attributes.get('role_name') == 'hero':         vehicle = actor         break if not vehicle:     raise RuntimeError("Ego vehicle with role_name 'hero' not found.")   # === NPC 차량 및 보행자 스폰 === spawn_points = world.get_map().get_spawn_points() random.shuffle(spawn_points)   npc_list = [] # 차량 스폰 for sp in spawn_points[:20]:     npc_bp = random.choice(bp_lib.filter('vehicle.*'))     npc = world.try_spawn_actor(npc_bp, sp)     if npc:         npc.set_autopilot(True)         npc_list.append(npc)         logging.info(f"Spawned NPC vehicle {npc.type_id} at {sp.location}")   # 보행자 스폰 pedestrian_list = [] controller_list = [] for sp in spawn_points[20:30]:  # 10명의 보행자     ped_bp = random.choice(bp_lib.filter('walker.pedestrian.*'))     try:         ped = world.try_spawn_actor(ped_bp, sp)         if ped:             pedestrian_list.append(ped)             controller_bp = bp_lib.find('controller.ai.walker')             controller = world.try_spawn_actor(controller_bp, carla.Transform(), attach_to=ped)             if controller:                 controller.start()                 controller.go_to_location(world.get_random_location_from_navigation())                 controller_list.append(controller)                 logging.info(f"Spawned pedestrian {ped.type_id} at {sp.location}")     except RuntimeError as e:         logging.warning(f"Pedestrian spawn failed: {e}")   # === 센서 블루프린트 === rgb_bp = bp_lib.find('sensor.camera.rgb') rgb_bp.set_attribute("image_size_x", str(IM_WIDTH)) rgb_bp.set_attribute("image_size_y", str(IM_HEIGHT)) rgb_bp.set_attribute("fov", "90") rgb_bp.set_attribute("sensor_tick", "0.05")  # 20fps   seg_bp = bp_lib.find('sensor.camera.semantic_segmentation') seg_bp.set_attribute("image_size_x", str(IM_WIDTH)) seg_bp.set_attribute("image_size_y", str(IM_HEIGHT)) seg_bp.set_attribute("fov", "90") seg_bp.set_attribute("sensor_tick", "0.05")  # 20fps   # === 센서 스폰 === cam_transform = carla.Transform(carla.Location(x=0.8, z=1.7)) rgb_cam = world.spawn_actor(rgb_bp, cam_transform, attach_to=vehicle) seg_cam = world.spawn_actor(seg_bp, cam_transform, attach_to=vehicle)   # === 동기 모드 설정 === settings = world.get_settings() settings.synchronous_mode = True settings.fixed_delta_seconds = 0.05 world.apply_settings(settings)   # === 큐 === rgb_q = queue.Queue() seg_q = queue.Queue() rgb_cam.listen(rgb_q.put) seg_cam.listen(seg_q.put)   # === Helper 함수들 === def build_K(w, h, fov):     f = w / (2.0 * np.tan(fov * np.pi / 360))     K = np.identity(3)     K[0, 0] = K[1, 1] = f     K[0, 2] = w / 2.0     K[1, 2] = h / 2.0     return K   def get_image_point(loc, K, w2c):     p = np.array([loc.x, loc.y, loc.z, 1.0])     pc = np.dot(w2c, p)     pc = [pc[1], -pc[2], pc[0]]     if pc[2] <= 0:         return None     pp = np.dot(K, pc)     pp[0] /= pp[2]     pp[1] /= pp[2]     return pp[0:2]   def point_in_image(p, h, w):     return p is not None and 0 <= p[0] < w and 0 <= p[1] < h   # === 메인 루프 === image_id = 0 last_save = time.time() K = build_K(IM_WIDTH, IM_HEIGHT, float(rgb_bp.get_attribute("fov")))   try:     while True:         world.tick()           rgb_img = rgb_q.get()         seg_img = seg_q.get()           # 카메라 pose         cam_tf = rgb_cam.get_transform()         w2c = np.array(cam_tf.get_inverse_matrix())           # RGB image         rgb = np.frombuffer(rgb_img.raw_data, dtype=np.uint8).reshape((IM_HEIGHT, IM_WIDTH, 4))         rgb = rgb[:, :, :3][:, :, ::-1].copy()  # RGB         rgb = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)         vis_img = rgb.copy()           # segmentation class ID (R 채널만)         seg_raw = np.frombuffer(seg_img.raw_data, dtype=np.uint8).reshape((IM_HEIGHT, IM_WIDTH, 4))         class_map = seg_raw[:, :, 2]  # Red = class ID           yolo_lines = []         coco_annotations = []           # 차량 및 보행자 처리         for actor in list(world.get_actors().filter("vehicle.*")) +              list(world.get_actors().filter("walker.pedestrian.*")):             if actor.id == vehicle.id:                 continue               dist = actor.get_transform().location.distance(vehicle.get_transform().location)             if dist > 70:                 continue               fwd = vehicle.get_transform().get_forward_vector()             rel = actor.get_transform().location - vehicle.get_transform().location             if fwd.dot(rel) < 0:                 continue               verts = actor.bounding_box.get_world_vertices(actor.get_transform())             x_min, y_min = IM_WIDTH, IM_HEIGHT             x_max, y_max = 0, 0             valid = False               for v in verts:                 p = get_image_point(v, K, w2c)                 if point_in_image(p, IM_HEIGHT, IM_WIDTH):                     x, y = int(p[0]), int(p[1])                     x_min = min(x_min, x)                     y_min = min(y_min, y)                     x_max = max(x_max, x)                     y_max = max(y_max, y)                     valid = True               if not valid or x_max <= x_min or y_max <= y_min:                 continue               crop = class_map[y_min:y_max, x_min:x_max]             if crop.size == 0:                 continue               # 클래스 ID 확인             class_name = 'pedestrian' if actor.type_id.startswith('walker.pedestrian') else 'vehicle'             carla_ids = CLASS_IDS[class_name]['carla_ids']             yolo_class_id = CLASS_IDS[class_name]['yolo_id']               visible_mask = np.isin(crop, carla_ids)             visible_ratio = np.sum(visible_mask) / (crop.shape[0] * crop.shape[1])             if visible_ratio < VISIBLE_THRESHOLD:                 continue               # YOLO format             x_center = ((x_min + x_max) / 2) / IM_WIDTH             y_center = ((y_min + y_max) / 2) / IM_HEIGHT             w_norm = (x_max - x_min) / IM_WIDTH             h_norm = (y_max - y_min) / IM_HEIGHT               yolo_lines.append(f"{yolo_class_id} {x_center:.6f} {y_center:.6f} {w_norm:.6f} {h_norm:.6f}")             cv2.rectangle(vis_img, (x_min, y_min), (x_max, y_max), (0, 255, 0), 2)             cv2.putText(vis_img, f"{class_name}: {visible_ratio*100:.1f}%", (x_min, y_min - 5),                          cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 0, 0), 2)               # COCO format             coco_annotations.append({                 "id": annotation_id,                 "image_id": image_id,                 "category_id": yolo_class_id,                 "bbox": [float(x_min), float(y_min), float(x_max - x_min), float(y_max - y_min)],                 "area": float((x_max - x_min) * (y_max - y_min)),                 "iscrowd": 0             })             annotation_id += 1           now = time.time()         if now - last_save >= SAVE_INTERVAL:             img_name = f"{image_id:06d}"             cv2.imwrite(os.path.join(OUTPUT_DIR, "images", f"{img_name}.jpg"), rgb)             with open(os.path.join(OUTPUT_DIR, "labels", f"{img_name}.txt"), "w") as f:                 f.writelines("\n".join(yolo_lines) + "\n")             logging.info(f"Saved image & labels: {img_name}")               # COCO JSON 업데이트             coco_json["images"].append({                 "id": image_id,                 "file_name": f"{img_name}.jpg",                 "width": IM_WIDTH,                 "height": IM_HEIGHT             })             coco_json["annotations"].extend(coco_annotations)                          image_id += 1             last_save = now           cv2.imshow("View", vis_img)         if cv2.waitKey(1) == ord('q'):             break   finally:     # COCO JSON 저장     try:         with open(os.path.join(COCO_DIR, "annotations.json"), 'w') as f:             json.dump(coco_json, f, indent=2)         logging.info("COCO annotations saved")     except Exception as e:         logging.error(f"Failed to save COCO annotations: {e}")       # 리소스 정리     rgb_cam.stop()     seg_cam.stop()     rgb_cam.destroy()     seg_cam.destroy()     for npc in npc_list:         npc.destroy()     for ped in pedestrian_list:         ped.destroy()     for controller in controller_list:         controller.stop()         controller.destroy()     world.apply_settings(carla.WorldSettings(synchronous_mode=False))     cv2.destroyAllWindows()     logging.info("Cleanup completed") Colored by Color Scripter

cs

데이터 확인

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

import os import cv2 import numpy as np import logging   # 로깅 설정 logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')   # 클래스 ID와 이름 매핑 CLASS_IDS = {     0: 'vehicle',     1: 'pedestrian' }   # 데이터셋 디렉토리 설정 dataset_dir = "./yolo_dataset" image_dir = os.path.join(dataset_dir, "images") label_dir = os.path.join(dataset_dir, "labels")   def load_yolo_labels(label_path, img_width, img_height):     """     YOLO 레이블 파일을 읽어 바운딩 박스 정보를 반환     """     annotations = []     try:         with open(label_path, 'r') as f:             lines = f.readlines()         for line in lines:             parts = line.strip().split()             if len(parts) != 5:                 logging.warning(f"잘못된 레이블 형식: {line.strip()} in {label_path}")                 continue             try:                 class_id = int(parts[0])                 x_center = float(parts[1]) * img_width                 y_center = float(parts[2]) * img_height                 width = float(parts[3]) * img_width                 height = float(parts[4]) * img_height                 # 픽셀 좌표로 변환                 x_min = int(x_center - width / 2)                 y_min = int(y_center - height / 2)                 x_max = int(x_center + width / 2)                 y_max = int(y_center + height / 2)                 annotations.append({                     'class_id': class_id,                     'bbox': [x_min, y_min, x_max, y_max]                 })             except ValueError as e:                 logging.warning(f"레이블 파싱 오류: {line.strip()} in {label_path}, {e}")                 continue         logging.info(f"YOLO 레이블 로드 완료: {label_path}")         return annotations     except Exception as e:         logging.error(f"레이블 파일 로드 실패: {label_path}, {e}")         return []   def draw_bounding_boxes(image, annotations):     """     이미지에 바운딩 박스를 그림     """     for ann in annotations:         class_id = ann['class_id']         x_min, y_min, x_max, y_max = [int(v) for v in ann['bbox']]         # 유효성 검사         if x_min < 0 or y_min < 0 or x_max > image.shape[1] or y_max > image.shape[0]:             logging.warning(f"유효하지 않은 바운딩 박스 좌표: {ann['bbox']}                              (이미지 크기: {image.shape[1]}x{image.shape[0]})")             continue         class_name = CLASS_IDS.get(class_id, 'unknown')         # 바운딩 박스 그리기         cv2.rectangle(image, (x_min, y_min), (x_max, y_max), (0, 0, 255), 1)         cv2.putText(image, class_name, (x_min, y_min - 10),                      cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1)     return image   def visualize_yolo_dataset():     """     YOLO 데이터셋을 시각화     """     if not os.path.exists(image_dir) or not os.path.exists(label_dir):         logging.error(f"이미지 또는 레이블 디렉토리가 존재하지 않습니다: {image_dir}, {label_dir}")         raise FileNotFoundError(f"이미지 또는 레이블 디렉토리가 존재하지 않습니다")       # 이미지 파일 목록 가져오기     image_files = sorted([f for f in os.listdir(image_dir) if f.endswith('.jpg')])       for image_filename in image_files:         image_path = os.path.join(image_dir, image_filename)         label_filename = image_filename.replace('.jpg', '.txt')         label_path = os.path.join(label_dir, label_filename)           # 이미지 로드         try:             image = cv2.imread(image_path)             if image is None:                 logging.error(f"이미지 로드 실패: {image_path}")                 continue         except Exception as e:             logging.error(f"이미지 로드 실패: {image_path}, {e}")             continue           img_height, img_width = image.shape[:2]           # 레이블 로드         if not os.path.exists(label_path):             logging.warning(f"레이블 파일이 존재하지 않습니다: {label_path}")             image_with_bboxes = image.copy()         else:             annotations = load_yolo_labels(label_path, img_width, img_height)             image_with_bboxes = draw_bounding_boxes(image.copy(), annotations)           # 이미지 표시         cv2.imshow("YOLO Dataset Visualization", image_with_bboxes)         logging.info(f"이미지 표시: {image_filename}")           while True:             key = cv2.waitKey(0) & 0xFF             if key == ord('n'):  # 'n' 키: 다음 이미지                 break             elif key == ord('q'):  # 'q' 키: 종료                 cv2.destroyAllWindows()                 logging.info("시각화 종료")                 return       cv2.destroyAllWindows()     logging.info("모든 이미지를 표시 완료")   if __name__ == "__main__":     try:         visualize_yolo_dataset()     except Exception as e:         logging.error(f"시각화 중 오류 발생: {e}")         cv2.destroyAllWindows() Colored by Color Scripter

cs

LineLane

데이터 수집 (차선, 주행 영역 하나의 코드에서 진행-CLASS_ID, OUTPUT_DIR, mask 수정 필요)

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

import sys import glob import numpy as np import cv2 from PIL import Image import os import time   # CARLA Python API 경로 설정 try:     sys.path.append(glob.glob('../carla/dist/carla-*%d.%d-%s.egg' % (         sys.version_info.major,         sys.version_info.minor,         'win-amd64' if os.name == 'nt' else 'linux-x86_64'))[0]) except IndexError:     pass   import carla   # =========================== # 설정 # =========================== # carla 클래스 # LANE_CLASS_ID = 24  # 차선 경계선 class (Carla semantic segmentation 기준) # 14번 차량 # 15번 트럭 # 18번 오토바이 # 20번 잡동사니 건물..? # 21번 잡동사니 # 1번 차도 # 2번 인도 # 3번 배경 # 6번 가로등 # 7번 신호등 엑터 # 8번 TrafficSign # 12번 사람. LANE_CLASS_ID = 24 # 차선 경계선 class (Carla semantic segmentation 기준) SAVE_DURATION = 120  # 저장할 시간(초) IMAGE_WIDTH = 1280 IMAGE_HEIGHT = 720 # OUTPUT_DIR = 'lane_dataset'  # 저장 경로 OUTPUT_DIR = 'drivable_dataset'  # 저장 경로   # =========================== # 센서 설정 # =========================== def spawn_segmentation_camera(world, vehicle):     bp_lib = world.get_blueprint_library()     cam_bp = bp_lib.find("sensor.camera.semantic_segmentation")     cam_bp.set_attribute("image_size_x", str(IMAGE_WIDTH))     cam_bp.set_attribute("image_size_y", str(IMAGE_HEIGHT))     cam_bp.set_attribute("fov", "90")     cam_bp.set_attribute("sensor_tick", "0.2")  # 0.2초마다 프레임 생성 (5fps)       transform = carla.Transform(carla.Location(x=1.5, z=2.4))     camera = world.spawn_actor(cam_bp, transform, attach_to=vehicle)     return camera   def spawn_traffic(world, num_vehicles=20):     blueprint_library = world.get_blueprint_library()     vehicle_blueprints = blueprint_library.filter('vehicle.*')       spawn_points = world.get_map().get_spawn_points()     if len(spawn_points) < num_vehicles:         num_vehicles = len(spawn_points)       vehicles = []     for i in range(num_vehicles):         bp = np.random.choice(vehicle_blueprints)         transform = spawn_points[i]         npc = world.try_spawn_actor(bp, transform)         if npc is not None:             npc.set_autopilot(True)             vehicles.append(npc)     print(f"🚗 총 {len(vehicles)} 대의 AI 차량이 스폰되었습니다.")     return vehicles # =========================== # YOLO Segmentation 저장 # =========================== def process_and_save(image, frame_id):     # Carla에서 받은 segmentation 이미지를 numpy 배열로 변환     array = np.frombuffer(image.raw_data, dtype=np.uint8)     array = array.reshape((IMAGE_HEIGHT, IMAGE_WIDTH, 4))[:, :, :3]     segmentation_mask = array[:, :, 2]  # Red 채널이 class ID         # 차선(RoadLines)만 추출     # mask = (segmentation_mask == LANE_CLASS_ID).astype(np.uint8) * 255     # 차선 + 도로 추출     INCLUDED_CLASSES = [1, 24]     mask = np.isin(segmentation_mask, INCLUDED_CLASSES).astype(np.uint8) * 255       # 윤곽선 찾기     contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)       label_lines = []     for contour in contours:         if len(contour) < 6:             continue  # 너무 작은 contour는 무시         # 정규화 (YOLO 포맷: 0~1 좌표)         normalized = [(pt[0][0] / IMAGE_WIDTH, pt[0][1] / IMAGE_HEIGHT) for pt in contour]         flattened = ' '.join([f"{x:.6f} {y:.6f}" for x, y in normalized])         label_lines.append(f"0 {flattened}")  # class 0 (lane)       # 이미지 저장 (RGB 변환)     rgb_image = Image.fromarray(array).convert("RGB")     img_path = os.path.join(OUTPUT_DIR, 'images', f"{frame_id:06d}.jpg")     label_path = os.path.join(OUTPUT_DIR, 'labels', f"{frame_id:06d}.txt")       rgb_image.save(img_path)       with open(label_path, 'w') as f:         f.write('\n'.join(label_lines))          # ------------------------     # 시각화: 원본 이미지 위에 차선 마스크 반투명 오버레이     # ------------------------     # OpenCV용 BGR 이미지로 변환     bgr_image = cv2.cvtColor(np.array(rgb_image), cv2.COLOR_RGB2BGR)       # 마스크 컬러 (녹색)     color_mask = np.zeros_like(bgr_image)     color_mask[mask == 255] = (0, 255, 0)  # 녹색       # 반투명 합성 (원본 70%, 마스크 30%)     overlayed = cv2.addWeighted(bgr_image, 0.7, color_mask, 0.3, 0)       # OpenCV 윈도우에 출력     cv2.imshow('Segmentation Overlay', overlayed)     cv2.waitKey(1)  # 1ms 대기 (필수, 그래야 이미지가 출력됨)       # =========================== # 메인 실행 # =========================== def main():     # Carla 클라이언트 연결     client = carla.Client('localhost', 2000)     client.set_timeout(5.0)     world = client.get_world()       # 차량 선택     vehicles = world.get_actors().filter('vehicle.*')     if not vehicles:         print("🚗 차량이 없습니다. 하나 스폰해 주세요.")         return     ego_vehicle = vehicles[0]       # 추가: 트래픽 생성     traffic_vehicles = spawn_traffic(world, num_vehicles=30)       # 저장 폴더 생성     os.makedirs(os.path.join(OUTPUT_DIR, 'images'), exist_ok=True)     os.makedirs(os.path.join(OUTPUT_DIR, 'labels'), exist_ok=True)       # 센서 설정     camera = spawn_segmentation_camera(world, ego_vehicle)       # 프레임 저장 루프     frame = {'id': 0}     def callback(image):         process_and_save(image, frame['id'])         print(f"📸 저장됨: frame {frame['id']}")         frame['id'] += 1       camera.listen(callback)       try:         print("🟢 차선 데이터 수집 시작...")         time.sleep(SAVE_DURATION)     finally:         camera.stop()         camera.destroy()         # 추가: 트래픽 차량 정리         for v in traffic_vehicles:             v.destroy()         print("🛑 수집 종료 및 모든 차량 제거 완료.")     if __name__ == '__main__':     main()   Colored by Color Scripter

cs

Drivable Area

주행 가능 영역 YOLO segment data

LineLane

차선 YOLO segment data

이미지 시각화 코드

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

import os import numpy as np import cv2   # =========================== # 설정 # =========================== # DATASET_DIR = 'lane_dataset'  # 이미지 및 라벨이 들어있는 상위 폴더 DATASET_DIR = 'drivable_dataset'  # 이미지 및 라벨이 들어있는 상위 폴더 IMG_DIR = os.path.join(DATASET_DIR, 'images') LBL_DIR = os.path.join(DATASET_DIR, 'labels') SAVE_DIR = os.path.join(DATASET_DIR, 'vis')  # 시각화 결과 저장 경로 os.makedirs(SAVE_DIR, exist_ok=True)   IMAGE_WIDTH = 1280 IMAGE_HEIGHT = 720   # =========================== # 시각화 루프 # =========================== file_list = sorted(os.listdir(LBL_DIR))   for label_file in file_list:     label_path = os.path.join(LBL_DIR, label_file)     img_file = label_file.replace('.txt', '.jpg')     img_path = os.path.join(IMG_DIR, img_file)       if not os.path.exists(img_path):         print(f"❌ 이미지 누락: {img_path}")         continue       # 이미지 로드     image = cv2.imread(img_path)       with open(label_path, 'r') as f:         lines = f.readlines()       for line in lines:         parts = line.strip().split()         if len(parts) < 3:             continue  # polygon이 아님           cls = int(parts[0])  # class id (0)         coords = list(map(float, parts[1:]))           # (x, y)로 좌표 쌍 묶기         pts = [(int(x * IMAGE_WIDTH), int(y * IMAGE_HEIGHT)) for x, y in zip(coords[::2], coords[1::2])]           if len(pts) >= 3:             cv2.polylines(image, [np.array(pts, dtype=np.int32)], isClosed=True, color=(0, 255, 0), thickness=2)       # 저장     out_path = os.path.join(SAVE_DIR, img_file)     cv2.imwrite(out_path, image)     print(f"✅ 시각화 저장됨: {out_path}")   Colored by Color Scripter

cs

데이터셋 형태.

TrafficLight

dataset 내부에 annotations.json 파일 존재

TrafficObject

coco_dataset 내부에 annotations.json파일 존재

0은 차량 1은 사람 객체

같은 클래스의 객체에 대해서 가려지는 경우에는 BBox가 생기는 문제 존재.

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

import os import cv2 import logging   # 종료 : q # 이미지 다음 : d, 이전 : a # BBox 다음 : s, 이전 : w # BBox 삭제 : n, 이미지, 라벨 삭제 : m   # 로깅 설정 logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')   IMAGE_DIR = 'yolo_dataset/images' LABEL_DIR = 'yolo_dataset/labels' IMAGE_EXT = '.jpg' LABEL_EXT = '.txt' WINDOW_NAME = 'BBox Editor'   image_files = sorted([f for f in os.listdir(IMAGE_DIR) if f.endswith(IMAGE_EXT)]) current_image_idx = 0 current_bbox_idx = 0   def load_image(image_path):     img = cv2.imread(image_path)     if img is None:         logging.error(f"이미지 로드 실패: {image_path}")         return None     return img   def load_labels(label_path):     if not os.path.exists(label_path):         logging.warning(f"레이블 파일이 존재하지 않습니다: {label_path}")         return []     try:         with open(label_path, 'r') as f:             return [line.strip() for line in f.readlines()]     except Exception as e:         logging.error(f"레이블 파일 로드 실패: {label_path}, {e}")         return []   def save_labels(label_path, labels):     try:         with open(label_path, 'w') as f:             f.write('\n'.join(labels) + '\n')         logging.info(f"레이블 저장 완료: {label_path}")     except Exception as e:         logging.error(f"레이블 저장 실패: {label_path}, {e}")   def draw_bbox(image, label_line, bbox_index, total_bboxes):     h, w = image.shape[:2]     parts = label_line.split()     if len(parts) != 5:         logging.warning(f"잘못된 레이블 형식: {label_line}")         return image     try:         cls, cx, cy, bw, bh = map(float, parts)         x1 = int((cx - bw / 2) * w)         y1 = int((cy - bh / 2) * h)         x2 = int((cx + bw / 2) * w)         y2 = int((cy + bh / 2) * h)         if x1 < 0 or y1 < 0 or x2 > w or y2 > h:             logging.warning(f"유효하지 않은 BBox 좌표: {label_line} (이미지 크기: {w}x{h})")             return image         img = image.copy()         cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)         cv2.putText(img, f"BBox {bbox_index + 1}/{total_bboxes}", (x1, y1 - 10),                     cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 2)         return img     except ValueError as e:         logging.warning(f"레이블 파싱 오류: {label_line}, {e}")         return image   def delete_image_and_label(image_path, label_path, image_name):     """     이미지와 레이블 파일을 데이터셋에서 삭제     """     try:         if os.path.exists(image_path):             os.remove(image_path)             logging.info(f"이미지 삭제 완료: {image_path}")         else:             logging.warning(f"이미지 파일이 존재하지 않습니다: {image_path}")         if os.path.exists(label_path):             os.remove(label_path)             logging.info(f"레이블 삭제 완료: {label_path}")         else:             logging.warning(f"레이블 파일이 존재하지 않습니다: {label_path}")         if image_name in image_files:             image_files.remove(image_name)             logging.info(f"이미지 목록에서 제거: {image_name}")         return True     except Exception as e:         logging.error(f"파일 삭제 실패: {image_path}, {label_path}, {e}")         return False   while True:     if not image_files:         logging.info("데이터셋에 이미지가 없습니다. 프로그램 종료.")         break       if current_image_idx < 0:         current_image_idx = 0     elif current_image_idx >= len(image_files):         current_image_idx = len(image_files) - 1       image_name = image_files[current_image_idx]     label_name = image_name.replace(IMAGE_EXT, LABEL_EXT)       img_path = os.path.join(IMAGE_DIR, image_name)     label_path = os.path.join(LABEL_DIR, label_name)       image = load_image(img_path)     if image is None:         logging.warning(f"이미지 로드 실패로 다음 이미지로 이동: {img_path}")         image_files.pop(current_image_idx)         if current_image_idx >= len(image_files):             current_image_idx = len(image_files) - 1         continue       labels = load_labels(label_path)       if current_bbox_idx >= len(labels):         current_bbox_idx = len(labels) - 1     if current_bbox_idx < 0:         current_bbox_idx = 0       display_img = image.copy()     if labels and current_bbox_idx < len(labels):         display_img = draw_bbox(display_img, labels[current_bbox_idx], current_bbox_idx, len(labels))       cv2.imshow(WINDOW_NAME, display_img)     key = cv2.waitKey(0) & 0xFF          if key == ord('q') or key == 27:  # 'q' 또는 ESC: 종료         break     elif key == ord('d'):  # 다음 이미지         current_image_idx += 1         current_bbox_idx = 0     elif key == ord('a'):  # 이전 이미지         current_image_idx -= 1         current_bbox_idx = 0     elif key == ord('s'):  # 다음 BBox         current_bbox_idx += 1         if current_bbox_idx >= len(labels):             current_bbox_idx = len(labels) - 1     elif key == ord('w'):  # 이전 BBox         current_bbox_idx -= 1         if current_bbox_idx < 0:             current_bbox_idx = 0     elif key == ord('n') and labels:  # 현재 BBox 삭제         logging.info(f"BBox {current_bbox_idx + 1} 삭제: {label_name}")         labels.pop(current_bbox_idx)         save_labels(label_path, labels)         if current_bbox_idx >= len(labels):             current_bbox_idx = len(labels) - 1     elif key == ord('m'):  # 이미지와 레이블 삭제         logging.info(f"데이터셋에서 삭제: {image_name}, {label_name}")         if delete_image_and_label(img_path, label_path, image_name):             if image_files:  # 리스트가 비어 있지 않으면 인덱스 조정                 if current_image_idx >= len(image_files):                     current_image_idx = len(image_files) - 1             else:                 break  # 이미지 목록이 비면 종료             current_bbox_idx = 0   cv2.destroyAllWindows() logging.info("프로그램 종료") Colored by Color Scripter

cs

다음 코드로 직접 BBox를 제거 및 해당 데이터를 제거하는 방법이 존재하나 차선책을 찾아야 함.