摘要：針對疫情期間存在的排查實時性差、排查效率低、無法追蹤密接者等問題，可以使用基於YOLOv4的行人檢測、行人距離估計、多目標跟蹤的方案進行解決。
本文分享自華為雲社群《基於ModelArts進行流感患者密接排查》，作者：HWCloudAI。
目前流感病毒患者密接難以排查，尤其是在人流量大的區域，進行排查需要消耗大量人力且需要等待。針對疫情期間存在的排查實時性差、排查效率低、無法追蹤密接者等問題，可以使用基於YOLOv4的行人檢測、行人距離估計、多目標跟蹤的方案進行解決。
1）利用行人重識別技術實現流感病毒患者及密接者識別功能；
2）結合Stereo-vision以及YOLO演算法實現患者的真實密切接觸鑑別；
3）利用SORT多目標跟蹤演算法繪製出患者及密接者的行動軌跡；
該系統可以有效提高防疫效率，減輕經濟與防疫壓力，提高安全性。
今天將帶大家瞭解 透過華為雲ModelArts的 DeepSocial-COVID-19社會距離監測案例實現AI排查新冠密接。
? 點選連結進入到AI Gallery的“DeepSocial-COVID-19社會距離監測”案例頁面，點選Run in ModelArts，即可進入ModelArts Jupyter執行環境，此處需要選用GPU的規格。
注：以下步驟所涉及的程式碼都已經寫好，直接點選程式碼前面的箭頭，讓其自動執行即可。
步驟一：從華為雲物件儲存服務（OBS）複製案例所需程式碼。
# 下載程式碼和資料 import moxing as mox mox.file.copy_parallel('obs://obs-aigallery-zc/clf/code/DeepSocial','DeepSocial') # 引入依賴 from IPython.display import display, Javascript, Image from base64 import b64decode, b64encode import os import cv2 import numpy as np import PIL import io import html import time import matplotlib.pyplot as plt %matplotlib inline
步驟二：在本地編譯YOLO。
需要根據執行環境修改Makefile 如是否有GPU等
如果編譯報錯：/bin/sh:nvcc not found
解決方式（參考）：
1）檢視nvcc可執行檔案的路徑
which nvcc
2）修改Makefile檔案中的NVCC=nvcc，把nvcc替換為上面查詢到的nvcc可執行檔案的路徑,如：/usr/local/cuda/bin/nvcc
NVCC=/usr/local/cuda/bin/nvcc
%cd DeepSocial
!make
步驟三：使用Darknet的python介面
# import darknet functions to perform object detections from darknet2 import * # load in our YOLOv4 architecture network network, class_names, class_colors = load_network("cfg/yolov4.cfg", "cfg/coco.data", "DeepSocial.weights") width = network_width(network) height = network_height(network) # darknet helper function to run detection on image def darknet_helper(img, width, height): darknet_image = make_image(width, height, 3) img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) img_resized = cv2.resize(img_rgb, (width, height), interpolation=cv2.INTER_LINEAR) # get image ratios to convert bounding boxes to proper size img_height, img_width, _ = img.shape width_ratio = img_width/width height_ratio = img_height/height # run model on darknet style image to get detections copy_image_from_bytes(darknet_image, img_resized.tobytes()) detections = detect_image(network, class_names, darknet_image) free_image(darknet_image) return detections, width_ratio, height_ratio
步驟四：使用SORT來實時跟蹤目標
!pip install filterpy from sort import * mot_tracker = Sort(max_age=25, min_hits=4, iou_threshold=0.3)
步驟五：輸入設定
Input = "OxfordTownCentreDataset.avi" # 需要檢測的適配 ReductionFactor = 2 # 取樣因子 calibration = [[180,162],[618,0],[552,540],[682,464]] # 相機標定的引數
步驟六：DeepSocial引數設定和函式引入
from deepsocial import * ######################## Frame number StartFrom = 0 EndAt = 500 #-1 for the end of the video ######################## (0:OFF/ 1:ON) Outputs CouplesDetection = 1 # Enable Couple Detection DTC = 1 # Detection, Tracking and Couples SocialDistance = 1 CrowdMap = 1 # MoveMap = 0 # ViolationMap = 0 # RiskMap = 0 ######################## Units are Pixel ViolationDistForIndivisuals = 28 ViolationDistForCouples = 31 #### CircleradiusForIndivsual = 14 CircleradiusForCouples = 17 ######################## MembershipDistForCouples = (16 , 10) # (Forward, Behind) per Pixel MembershipTimeForCouples = 35 # Time for considering as a couple (per Frame) ######################## (0:OFF/ 1:ON) CorrectionShift = 1 # Ignore people in the margins of the video HumanHeightLimit = 200 # Ignore people with unusual heights ######################## Transparency = 0.7 ######################## Output Video's path Path_For_DTC = os.getcwd() + "/DeepSOCIAL DTC.mp4" Path_For_SocialDistance = os.getcwd() + "/DeepSOCIAL Social Distancing.mp4" Path_For_CrowdMap = os.getcwd() + "/DeepSOCIAL Crowd Map.mp4" def extract_humans(detections): detetcted = [] if len(detections) > 0: # At least 1 detection in the image and check detection presence in a frame idList = [] id = 0 for label, confidence, bbox in detections: if label == 'person': xmin, ymin, xmax, ymax = bbox2points(bbox) id +=1 if id not in idList: idList.append(id) detetcted.append([int(xmin), int(ymin), int(xmax), int(ymax), idList[-1]]) return np.array(detetcted) def centroid(detections, image, calibration, _centroid_dict, CorrectionShift, HumanHeightLimit): e = birds_eye(image.copy(), calibration) centroid_dict = dict() now_present = list() if len(detections) > 0: for d in detections: p = int(d[4]) now_present.append(p) xmin, ymin, xmax, ymax = d[0], d[1], d[2], d[3] w = xmax - xmin h = ymax - ymin x = xmin + w/2 y = ymax - h/2 if h < HumanHeightLimit: overley = e.image bird_x, bird_y = e.projection_on_bird((x, ymax)) if CorrectionShift: if checkupArea(overley, 1, 0.25, (x, ymin)): continue e.setImage(overley) center_bird_x, center_bird_y = e.projection_on_bird((x, ymin)) centroid_dict[p] = ( int(bird_x), int(bird_y), int(x), int(ymax), int(xmin), int(ymin), int(xmax), int(ymax), int(center_bird_x), int(center_bird_y)) _centroid_dict[p] = centroid_dict[p] return _centroid_dict, centroid_dict, e.image def ColorGenerator(seed=1, size=10): np.random.seed = seed color=dict() for i in range(size): h = int(np.random.uniform() *255) color[i]= h return color def VisualiseResult(_Map, e): Map = np.uint8(_Map) histMap = e.convrt2Image(Map) visualBird = cv2.applyColorMap(np.uint8(_Map), cv2.COLORMAP_JET) visualMap = e.convrt2Image(visualBird) visualShow = cv2.addWeighted(e.original, 0.7, visualMap, 1 - 0.7, 0) return visualShow, visualBird, histMap
步驟七：推理過程
cap = cv2.VideoCapture(Input) frame_width = int(cap.get(3)) frame_height = int(cap.get(4)) height, width = frame_height // ReductionFactor, frame_width // ReductionFactor print("Video Reolution: ",(width, height)) if DTC: DTCVid = cv2.VideoWriter(Path_For_DTC, cv2.VideoWriter_fourcc(*'X264'), 30.0, (width, height)) if SocialDistance: SDimageVid = cv2.VideoWriter(Path_For_SocialDistance, cv2.VideoWriter_fourcc(*'X264'), 30.0, (width, height)) if CrowdMap: CrowdVid = cv2.VideoWriter(Path_For_CrowdMap, cv2.VideoWriter_fourcc(*'X264'), 30.0, (width, height)) colorPool = ColorGenerator(size = 3000) _centroid_dict = dict() _numberOFpeople = list() _greenZone = list() _redZone = list() _yellowZone = list() _final_redZone = list() _relation = dict() _couples = dict() _trackMap = np.zeros((height, width, 3), dtype=np.uint8) _crowdMap = np.zeros((height, width), dtype=np.int) _allPeople = 0 _counter = 1 frame = 0 while True: print('-- Frame : {}'.format(frame)) prev_time = time.time() ret, frame_read = cap.read() if not ret: break frame += 1 if frame <= StartFrom: continue if frame != -1: if frame > EndAt: break frame_resized = cv2.resize(frame_read,(width, height), interpolation=cv2.INTER_LINEAR) image = frame_resized e = birds_eye(image, calibration) detections, width_ratio, height_ratio = darknet_helper(image, width, height) humans = extract_humans(detections) track_bbs_ids = mot_tracker.update(humans) if len(humans) != 0 else humans _centroid_dict, centroid_dict, partImage = centroid(track_bbs_ids, image, calibration, _centroid_dict, CorrectionShift, HumanHeightLimit) redZone, greenZone = find_zone(centroid_dict, _greenZone, _redZone, criteria=ViolationDistForIndivisuals) if CouplesDetection: _relation, relation = find_relation(e, centroid_dict, MembershipDistForCouples, redZone, _couples, _relation) _couples, couples, coupleZone = find_couples(image, _centroid_dict, relation, MembershipTimeForCouples, _couples) yellowZone, final_redZone, redGroups = find_redGroups(image, centroid_dict, calibration, ViolationDistForCouples, redZone, coupleZone, couples , _yellowZone, _final_redZone) else: couples = [] coupleZone = [] yellowZone = [] redGroups = redZone final_redZone = redZone if DTC: DTC_image = image.copy() _trackMap = Apply_trackmap(centroid_dict, _trackMap, colorPool, 3) DTC_image = cv2.add(e.convrt2Image(_trackMap), image) DTCShow = DTC_image for id, box in centroid_dict.items(): center_bird = box[0], box[1] if not id in coupleZone: cv2.rectangle(DTCShow,(box[4], box[5]),(box[6], box[7]),(0,255,0),2) cv2.rectangle(DTCShow,(box[4], box[5]-13),(box[4]+len(str(id))*10, box[5]),(0,200,255),-1) cv2.putText(DTCShow,str(id),(box[4]+2, box[5]-2),cv2.FONT_HERSHEY_SIMPLEX,.4,(0,0,0),1,cv2.LINE_AA) for coupled in couples: p1 , p2 = coupled couplesID = couples[coupled]['id'] couplesBox = couples[coupled]['box'] cv2.rectangle(DTCShow, couplesBox[2:4], couplesBox[4:], (0,150,255), 4) loc = couplesBox[0] , couplesBox[3] offset = len(str(couplesID)*5) captionBox = (loc[0] - offset, loc[1]-13), (loc[0] + offset, loc[1]) cv2.rectangle(DTCShow,captionBox[0],captionBox[1],(0,200,255),-1) wc = captionBox[1][0] - captionBox[0][0] hc = captionBox[1][1] - captionBox[0][1] cx = captionBox[0][0] + wc // 2 cy = captionBox[0][1] + hc // 2 textLoc = (cx - offset, cy + 4) cv2.putText(DTCShow, str(couplesID) ,(textLoc),cv2.FONT_HERSHEY_SIMPLEX,.4,(0,0,0),1,cv2.LINE_AA) DTCVid.write(DTCShow) if SocialDistance: SDimage, birdSDimage = Apply_ellipticBound(centroid_dict, image, calibration, redZone, greenZone, yellowZone, final_redZone, coupleZone, couples, CircleradiusForIndivsual, CircleradiusForCouples) SDimageVid.write(SDimage) if CrowdMap: _crowdMap, crowdMap = Apply_crowdMap(centroid_dict, image, _crowdMap) crowd = (crowdMap - crowdMap.min()) / (crowdMap.max() - crowdMap.min())*255 crowd_visualShow, crowd_visualBird, crowd_histMap = VisualiseResult(crowd, e) CrowdVid.write(crowd_visualShow) cv2.waitKey(3) print('::: Analysis Completed') cap.release() if DTC: DTCVid.release(); print("::: Video Write Completed : ", Path_For_DTC) if SocialDistance: SDimageVid.release() ; print("::: Video Write Completed : ", Path_For_SocialDistance) if CrowdMap: CrowdVid.release() ; print("::: Video Write Completed : ", Path_For_CrowdMap)
步驟八：展示結果
from IPython.display import HTML outpath = "DeepSOCIAL DTC.mp4" mp4 = open(outpath,'rb').read() data_url = "data:video/mp4;base64," + b64encode(mp4).decode() HTML(""" <video width=400 controls> <source src="%s" type="video/mp4"> </video> """ % data_url)
<iframe src="https://obs-aigallery-zc.obs.cn-north-4.myhuaweicloud.com/clf/code/DeepSocial/DeepSOCIAL%20DTC.mp4" scrolling="no" border="0" frameborder="no" framespacing="0" allowfullscreen="true" height=450 width=800> </iframe> <iframe src="https://obs-aigallery-zc.obs.cn-north-4.myhuaweicloud.com/clf/code/DeepSocial/DeepSOCIAL%20Social%20Distancing.mp4" scrolling="no" border="0" frameborder="no" framespacing="0" allowfullscreen="true" height=450 width=800> </iframe>
如果想要更好的效果，如何進行最佳化呢？
1.使用精確度更高的檢測演算法YOLOv7，使用追蹤效果更好的Deep SORT；
2.使用更多資料進行訓練
本次介紹就到這裡啦，大家快去Gallery實操一下吧！