ll =[]
	for b0 in range(0,len(output)):
		b0y,b0x,b0r = output[b0]
		cv2.circle(clone1, (int(b0x),int(b0y)), int(b0r), (0, 0, 250), 1)             
		l =[]
		kp_1 =[]
		ds_1 =[]
		l3 =[]
		index= 0
		for  k,d in zip(key,des):
			if (k.pt[0] - b0x)**2 + (k.pt[1] - b0y)**2  <= (b0r **2):
				l.append(index)
				#print('l :',len(l))
				kp_1.append(k)
				ds_1.append(d)
			index+=1
		if l:
			kp_2= np.delete(key,l,axis=0)
			ds_2 = np.delete(des,l,axis=0)
			#print('k :',len(kp),'...',len(ds))
			#nn_matches = bf.match(np.array(ds_1),ds_2)
			nn_matches = matcher.knnMatch(np.array(ds_1), ds_2, 2)
			#print(nn_matches)
			good = []
			#matched1 = []
			#matched2 = []
			nn_match_ratio = 0.6 # Nearest neighbor matching ratio
			for m, n in nn_matches:
				#print(m)
				#Use 2-nn matches and ratio criterion to find correct keypoint matches
				#If the closest match distance is significantly lower than the second closest one, then the match is correct (match is not ambiguous).
				if m.distance < nn_match_ratio * n.distance:
					#print(x1,y1,x2,y2)
					good.append(m)


			MIN_MATCH_COUNT = 3
			if len(good) > MIN_MATCH_COUNT:
				src_pts = np.float32([kp_1[m.queryIdx].pt for m in good]).reshape(-1, 1, 2)
				dst_pts = np.float32([kp_2[m.trainIdx].pt for m in good]).reshape(-1, 1, 2)
				#src = np.concatenate((src,src_pts))
				#dst = np.concatenate((dst,dst_pts))
				src = np.array(src_pts).ravel()
				dst = np.array(dst_pts).ravel()
				ps =np.array(src).reshape((-1,2))
				pd =np.array(dst).reshape((-1,2))
				for k1,k2 in zip(ps,pd):
					cv2.circle(clone1, (int(k1[0]),int(k1[1])),4,(0,0,255),-1)
					cv2.circle(clone1, (int(k2[0]),int(k2[1])),4,(0,255,255),-1)
					cv2.line(clone1,(int(k1[0]),int(k1[1])),(int(k2[0]),int(k2[1])),(0,255,0),2)           
	#cv2.imshow('image',clone1)
	cv2.imwrite('detectionz-results__'+str(i)+'.png',clone1)    
	end_time = datetime.now()
	print('Duration: {}'.format(end_time - start_time))
	i += 1
cv2.waitKey(0)
cv2.destroyAllWindows()