偵測攝影機中物體

import cv2

import numpy as np

net = cv2.dnn.readNetFromDarknet("cfg/yolov3.cfg","yolov3.weights") #讀取模型

layer_names = net.getLayerNames()

output_layers = [layer_names[i[0] -1] for i in net.getUnconnectedOutLayers()] #輸出圖形規格

classes = [line.strip() for line in open("data/coco.names")] #分類標籤

colors = [(255,0,0),(0,255,0),(0,0,255),(127,0,255),(0,125,255)] #框選顏色

cap = cv2.VideoCapture(0) #打開攝影機

while(True):

    ret , img = cap.read() #讀取影像

    height, width , channels = img.shape

    blob = cv2.dnn.blobFromImage(img , 1/255.0 , (416,416) , (0,0,0),True ,crop=False) #圖形預處理以符合輸入圖片規格

    net.setInput(blob)

    outs = net.forward(output_layers) #偵測結果

    class_ids = [] #存標籤索引

    confidences = [] #存信心指數

    boxes = [] #存矩形座標

    for out in outs:

        for detection in out:

            tx, ty, tw, th, confidence = detection[0:5] #取得座標及信心資料

            scores = detection[5:]

            class_id = np.argmax(scores)

            if confidence >0.3: #信心指數大於0.3才算

                center_x = int(tx * width)

                center_y = int(ty * height)

                w = int(tw * width)

                h = int(th * height)

                #計算矩形方框資訊

                x = int(center_x -w /2)

                y = int(center_y -h /2)

                boxes.append([x,y,w,h])

                confidences.append(float(confidence))

                class_ids.append(class_id)

    indexes = cv2.dnn.NMSBoxes(boxes , confidences,0.3,0.4) #消除重疊框選

    font = cv2.FONT_HERSHEY_PLAIN

    for j in range(len(boxes)):

        if j in indexes:

            x,y,w,h = boxes[j]

            label = str(classes[class_ids[j]])

            color = colors[class_ids[i]%5]

            cv2.rectangle(img , (x,y),(x+w,y+h),color,1)

            cv2.putText(img , label,(x,y-5),font,1,color,2)

    cv2.imshow(f'1',img)

    if cv2.waitKey(10) == 27: #按ESC鍵結束迴圈

        break

cap.release()

cv2.destroyAllWindows()

Python程式進行攝影機偵測物體，是以偵測圖片中物體為基礎，對攝影機的每一幀圖片進行偵測，所以是不斷擷取攝影機畫面進行處理，偵測單張圖片物體並顯示。