k20hcmus/FishEye8K
3
1import argparse2import os3import platform4import sys5from pathlib import Path6import math7import torch8import numpy as np9import re10from deep_sort_pytorch.utils.parser import get_config11from deep_sort_pytorch.deep_sort import DeepSort12import pandas as pd13from collections import deque14FILE = Path(__file__).resolve()15ROOT = FILE.parents[0] # YOLO root directory16if str(ROOT) not in sys.path:17 sys.path.append(str(ROOT)) # add ROOT to PATH18ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative19 20from models.common import DetectMultiBackend21from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadScreenshots, LoadStreams22from utils.general import (LOGGER, Profile, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2,23 increment_path, non_max_suppression, print_args, scale_boxes, strip_optimizer, xyxy2xywh)24from utils.plots import Annotator, colors, save_one_box25from utils.torch_utils import select_device, smart_inference_mode26 27def initialize_deepsort():28 # Create the Deep SORT configuration object and load settings from the YAML file29 cfg_deep = get_config()30 cfg_deep.merge_from_file("deep_sort_pytorch/configs/deep_sort.yaml")31 32 # Initialize the DeepSort tracker33 deepsort = DeepSort(cfg_deep.DEEPSORT.REID_CKPT,34 max_dist=cfg_deep.DEEPSORT.MAX_DIST,35 # min_confidence parameter sets the minimum tracking confidence required for an object detection to be considered in the tracking process36 min_confidence=cfg_deep.DEEPSORT.MIN_CONFIDENCE,37 #nms_max_overlap specifies the maximum allowed overlap between bounding boxes during non-maximum suppression (NMS)38 nms_max_overlap=cfg_deep.DEEPSORT.NMS_MAX_OVERLAP,39 #max_iou_distance parameter defines the maximum intersection-over-union (IoU) distance between object detections40 max_iou_distance=cfg_deep.DEEPSORT.MAX_IOU_DISTANCE,41 # Max_age: If an object's tracking ID is lost (i.e., the object is no longer detected), this parameter determines how many frames the tracker should wait before assigning a new id42 max_age=cfg_deep.DEEPSORT.MAX_AGE, n_init=cfg_deep.DEEPSORT.N_INIT,43 #nn_budget: It sets the budget for the nearest-neighbor search.44 nn_budget=cfg_deep.DEEPSORT.NN_BUDGET,45 use_cuda=False46 )47 48 return deepsort49 50deepsort = initialize_deepsort()51data_deque = {}52def classNames():53 cocoClassNames = ["Bus", "Bike", "Car", "Pedestrian", "Truck"54 ]55 return cocoClassNames56className = classNames()57# def convert_to_int(x):58# if isinstance(x, str):59# # Extract numeric value from tensor string using regular expressions60# match = re.match(r'tensor\((\d+)\)', x)61# if match:62# return int(match.group(1))63# return x64def colorLabels(classid):65 if classid == 0: #Bus66 color = (0, 0, 255)67 elif classid == 1: #Bike 250, 247, 068 color = (255, 148, 0) # BGR (247, 0, 250)69 elif classid == 2: #Car 70 color = (0, 255, 10)71 elif classid == 3: #Pedestrian72 color = (0, 247, 250) 73 else: #Truck74 color = (235,0,255) 75 return tuple(color)76 77def convert_to_int(tensor):78 return tensor.type(torch.int16).item()79 80def draw_boxes(frame, bbox_xyxy, draw_trails, identities=None, categories=None, offset=(0,0)):81 height, width, _ = frame.shape82 for key in list(data_deque):83 if key not in identities:84 data_deque.pop(key)85 86 for i, box in enumerate(bbox_xyxy):87 x1, y1, x2, y2 = [int(i) for i in box]88 x1 += offset[0]89 y1 += offset[0]90 x2 += offset[0]91 y2 += offset[0]92 #Find the center point of the bounding box93 center = int((x1+x2)/2), int((y1+y2)/2)94 cat = int(categories[i]) if categories is not None else 095 color = colorLabels(cat)96 #color = [255,0,0]#compute_color_labels(cat)97 id = int(identities[i]) if identities is not None else 098 # create new buffer for new object99 if id not in data_deque:100 data_deque[id] = deque(maxlen= 64)101 data_deque[id].appendleft(center)102 cv2.rectangle(frame, (x1, y1), (x2, y2), color, 2)103 # name = className[cat]104 # label = str(id) + ":" + name105 # text_size = cv2.getTextSize(label, 0, fontScale=0.5, thickness=2)[0]106 # c2 = x1 + text_size[0], y1 - text_size[1] - 3107 # cv2.rectangle(frame, (x1, y1), c2, color, -1)108 # cv2.putText(frame, label, (x1, y1 - 2), 0, 0.5, [255, 255, 255], thickness=1, lineType=cv2.LINE_AA)109 cv2.circle(frame,center, 2, (0,255,0), cv2.FILLED)110 if draw_trails:111 # draw trail112 for i in range(1, len(data_deque[id])):113 # check if on buffer value is none114 if data_deque[id][i - 1] is None or data_deque[id][i] is None:115 continue116 # generate dynamic thickness of trails117 thickness = int(np.sqrt(64 / float(i + i)) * 1.5)118 # draw trails119 cv2.line(frame, data_deque[id][i - 1], data_deque[id][i], color, thickness) 120 return frame121 122@smart_inference_mode()123def run_deepsort(124 weights=ROOT / 'yolo.pt', # model path or triton URL125 source=ROOT / 'data/images', # file/dir/URL/glob/screen/0(webcam)126 data=ROOT / 'data/coco.yaml', # dataset.yaml path127 imgsz=(640, 640), # inference size (height, width)128 conf_thres=0.25, # confidence threshold129 iou_thres=0.45, # NMS IOU threshold130 max_det=1000, # maximum detections per image131 device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu132 view_img=False, # show results133 nosave=False, # do not save images/videos134 classes=None, # filter by class: --class 0, or --class 0 2 3135 agnostic_nms=False, # class-agnostic NMS136 augment=False, # augmented inference137 visualize=False, # visualize features138 update=False, # update all models139 project=ROOT / 'runs/detect', # save results to project/name140 name='exp', # save results to project/name141 exist_ok=False, # existing project/name ok, do not increment142 half=False, # use FP16 half-precision inference143 dnn=False, # use OpenCV DNN for ONNX inference144 vid_stride=1, # video frame-rate stride145 draw_trails = False,146):147 source = str(source)148 save_img = not nosave and not source.endswith('.txt') # save inference images149 is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)150 is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))151 webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file)152 screenshot = source.lower().startswith('screen')153 if is_url and is_file:154 source = check_file(source) # download155 156 # Directories157 save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run158 save_dir.mkdir(parents=True, exist_ok=True) # make dir159 160 # Load model161 device = select_device(device)162 model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)163 stride, names, pt = model.stride, model.names, model.pt164 imgsz = check_img_size(imgsz, s=stride) # check image size165 166 # Dataloader167 bs = 1 # batch_size168 if webcam:169 view_img = check_imshow(warn=True)170 dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)171 bs = len(dataset)172 elif screenshot:173 dataset = LoadScreenshots(source, img_size=imgsz, stride=stride, auto=pt)174 else:175 dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)176 vid_path, vid_writer = [None] * bs, [None] * bs177 178 # Run inference179 model.warmup(imgsz=(1 if pt or model.triton else bs, 3, *imgsz)) # warmup180 seen, windows, dt = 0, [], (Profile(), Profile(), Profile())181 frame_counts = [] 182 for path, im, im0s, vid_cap, s in dataset:183 with dt[0]:184 im = torch.from_numpy(im).to(model.device)185 im = im.half() if model.fp16 else im.float() # uint8 to fp16/32186 im /= 255 # 0 - 255 to 0.0 - 1.0187 if len(im.shape) == 3:188 im = im[None] # expand for batch dim189 190 # Inference191 with dt[1]:192 visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False193 pred = model(im, augment=augment, visualize=visualize)194 # pred = pred[0][1]195 196 # NMS197 with dt[2]:198 pred = pred[0][1] if isinstance(pred[0], list) else pred[0] # single model or ensemble199 pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det)200 201 # Second-stage classifier (optional)202 # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s)203 counts = {}204 # Process predictions205 for i, det in enumerate(pred): # per image206 seen += 1207 if webcam: # batch_size >= 1208 p, im0, frame = path[i], im0s[i].copy(), dataset.count209 s += f'{i}: '210 else:211 p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0)212 213 p = Path(p) # to Path214 save_path = str(save_dir / p.name) # im.jpg215 txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}') # im.txt216 s += '%gx%g ' % im.shape[2:] # print string217 gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh218 ims = im0.copy()219 if len(det):220 # Rescale boxes from img_size to im0 size221 det[:, :4] = scale_boxes(im.shape[2:], det[:, :4], im0.shape).round()222 223 # Print results224 for c in det[:, 5].unique():225 n = (det[:, 5] == c).sum() # detections per class226 s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string227 counts[names[int(c)]] = n228 xywh_bboxs = []229 confs = []230 oids = []231 outputs = []232 # Write results233 for *xyxy, conf, cls in reversed(det):234 x1, y1, x2, y2 = xyxy235 x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)236 #Find the Center Coordinates for each of the detected object237 cx, cy = int((x1+x2)/2), int((y1+y2)/2)238 #Find the Width and Height of the Boundng box239 bbox_width = abs(x1-x2)240 bbox_height = abs(y1-y2)241 xcycwh = [cx, cy, bbox_width, bbox_height]242 xywh_bboxs.append(xcycwh)243 conf = math.ceil(conf*100)/100244 confs.append(conf)245 classNameInt = int(cls)246 oids.append(classNameInt)247 xywhs = torch.tensor(xywh_bboxs)248 confss = torch.tensor(confs) 249 outputs = deepsort.update(xywhs, confss, oids, ims)250 if len(outputs) > 0:251 bbox_xyxy = outputs[:, :4]252 identities = outputs[:, -2]253 object_id = outputs[:, -1]254 draw_boxes(ims, bbox_xyxy, draw_trails, identities, object_id)255 256 # Stream results257 if view_img:258 if platform.system() == 'Linux' and p not in windows:259 windows.append(p)260 cv2.namedWindow(str(p), cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO) # allow window resize (Linux)261 cv2.resizeWindow(str(p), ims.shape[1], ims.shape[0])262 cv2.imshow(str(p), ims)263 cv2.waitKey(1) # 1 millisecond264 # Save results (image with detections)265 if save_img:266 if vid_path[i] != save_path: # new video267 vid_path[i] = save_path268 if isinstance(vid_writer[i], cv2.VideoWriter):269 vid_writer[i].release() # release previous video writer270 if vid_cap: # video271 fps = vid_cap.get(cv2.CAP_PROP_FPS)272 w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))273 h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))274 else: # stream275 fps, w, h = 30, ims.shape[1], ims.shape[0]276 save_path = str(Path(save_path).with_suffix('.mp4')) # force *.mp4 suffix on results videos277 vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc('m','p','4','v'), fps, (w, h))278 vid_writer[i].write(ims)279 280 # Print time (inference-only)281 LOGGER.info(f"{s}{'' if len(det) else '(no detections), '}{dt[1].dt * 1E3:.1f}ms")282 frame_counts.append((frame, counts)) # Append the counts for each frame283 transformed_data = []284 285 # Iterate over frame_counts and transform each entry into a row in the DataFrame286 for frame, counts_dict in frame_counts:287 for label, count in counts_dict.items():288 transformed_data.append((frame, label.capitalize(), count))289 290 # Create a DataFrame from the transformed data291 df = pd.DataFrame(transformed_data, columns=['frame', 'label', 'count'])292 293 # Convert count column from tensors to integers294 df['count'] = df['count'].apply(convert_to_int)295 296 counts_df = pd.DataFrame(counts.items(), columns=['label', 'count'])297 counts_df['count'] = counts_df['count'].apply(convert_to_int)298 counts_df['label'] = counts_df['label'].astype(str) 299 300 if update:301 strip_optimizer(weights[0]) # update model (to fix SourceChangeWarning)302 return save_path, counts_df, df 303 304 305def parse_opt():306 parser = argparse.ArgumentParser()307 parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo.pt', help='model path or triton URL')308 parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob/screen/0(webcam)')309 parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path')310 parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w')311 parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold')312 parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')313 parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image')314 parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')315 parser.add_argument('--view-img', action='store_true', help='show results')316 parser.add_argument('--nosave', action='store_true', help='do not save images/videos')317 parser.add_argument('--draw-trails', action='store_true', help='do not drawtrails')318 parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3')319 parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')320 parser.add_argument('--augment', action='store_true', help='augmented inference')321 parser.add_argument('--visualize', action='store_true', help='visualize features')322 parser.add_argument('--update', action='store_true', help='update all models')323 parser.add_argument('--project', default=ROOT / 'runs/detect', help='save results to project/name')324 parser.add_argument('--name', default='exp', help='save results to project/name')325 parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')326 parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')327 parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')328 parser.add_argument('--vid-stride', type=int, default=1, help='video frame-rate stride')329 opt = parser.parse_args()330 opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1 # expand331 print_args(vars(opt))332 return opt333 334 335def main(opt):336 # check_requirements(exclude=('tensorboard', 'thop'))337 run_deepsort(**vars(opt))338 339 340 341if __name__ == "__main__":342 opt = parse_opt()343 main(opt)