使用Paddlehub能够便捷地获取PaddlePaddle生态下的预训练模型，完成模型的管理和一键预测。配合使用Fine-tune API，可以基于大规模预训练模型快速完成迁移学习，让预训练模型能更好地服务于用户特定场景的应用。
本项目采用Paddlehub, 基于human_pose_estimation_resnet50_mpii人体关键点检测模型。
# 使用前先安装模型
hub install human_pose_estimation_resnet50_mpii==1.1.1
人体骨骼关键点检测(Pose Estimation) 是计算机视觉的基础性算法之一，在诸多计算机视觉任务起到了基础性的作用，如行为识别、人物跟踪、步态识别等相关领域。具体应用主要集中在智能视频监控，病人监护系统，人机交互，虚拟现实，人体动画，智能家居，智能安防，运动员辅助训练等等。 该模型的论文《Simple Baselines for Human Pose Estimation and Tracking》由 MSRA 发表于 ECCV18，使用 MPII 数据集训练完成。
本项目实现在健身过程的哑铃抬举的过程中，对进行此过程的六个关节点分别检测，即左右手腕、左右手肘、左右肩部，并分别计算以手肘为基点，手腕、手肘、肩部所构成的向量夹角。以夹角为核心 判断哑铃动作是否到位，完成两次的【0°-90°】向量夹角变化算作一次哑铃抬举动作。
Step one
使用human_pose_estimation_resnet50_mpii模型对受众人体关键点进行检测，并从中抽取出项目所需的六个关节点，即：【right_wrist、right_elbow、right_shoulder、left_wrist、left_elbow、left_shoulder】,并将其存放进列表中：
def aitrainer(picture):
consequence = []
result = pose_estimation.keypoint_detection(images=[cv2.imread(picture)])
predict = dict(result[0]["data"])
right_wrist = predict['right_wrist']
consequence.append(right_wrist)
right_elbow = predict['right_elbow']
consequence.append(right_elbow)
right_shoulder = predict['right_shoulder']
consequence.append(right_shoulder)
left_shoulder = predict['left_shoulder']
consequence.append(left_shoulder)
left_elbow = predict['left_elbow']
consequence.append(left_elbow)
left_wrist = predict['left_wrist']
consequence.append(left_wrist)
return consequence
Step two
前文提及，本项目的核心在于【以手肘为中心点，对手腕、手肘、肩部所构成的夹角进行计算，以夹角为判断依据，从而对哑铃抬举动作进行监督并且计数】。
def Cal_Ang(point_1, point_2, point_3):
"""
根据三点坐标计算夹角
:param point_1: 点1坐标
:param point_2: 点2坐标
:param point_3: 点3坐标
:return: 点2的夹角(中心点)
"""
a=math.sqrt((point_2[0]-point_3[0]) * (point_2[0]-point_3[0])+(point_2[1]-point_3[1]) * (point_2[1]-point_3[1]))
b=math.sqrt((point_1[0]-point_3[0]) * (point_1[0]-point_3[0])+(point_1[1]-point_3[1]) * (point_1[1]-point_3[1]))
c=math.sqrt((point_1[0]-point_2[0]) * (point_1[0]-point_2[0])+(point_1[1]-point_2[1]) * (point_1[1]-point_2[1]))
B=math.degrees(math.acos((b*b-a*a-c*c)/(-2*a*c)))
return B
Step three
使用OpenCV打开本地摄像头获取动态页面，并将项目检测结果实时展示在页面中：
cap = cv2.VideoCapture(0)
while(1):
# get a frame
ret, frame = cap.read()
frame = cv2.resize(frame, (1280, 720))
print("训练将在5秒后进行，请您做好准备！")
for i in range(5):
time.sleep(1)
m = 5 - i
print(m)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cv2.imwrite("camera.jpg", frame)
point = aitrainer('camera.jpg')
if train_content == '左臂':
left_wrist = point[5]
left_elbow = point[4]
left_shoulder = point[3]
angle_left = Cal_Ang(left_wrist,left_elbow,left_shoulder)
diff = angle_left - angle_last
if diff >=0 and det_1 == 0:
count += 0.5
det_1 = 1
if diff <=0 and det_1 ==1:
count += 0.5
det_1 = 2
cv2.line(frame, tuple(left_wrist), tuple(left_elbow),(0, 0, 255), 3)
cv2.line(frame, tuple(left_elbow), tuple(left_shoulder),(0, 0, 225), 3)
if count % 1.0 == 0.0:
det_1 = 0
print("count:%.2f"%count)
cv2.rectangle(frame, (0, 450), (250, 720), (0, 255, 0), cv2.FILLED)
cv2.putText(frame, str(int(count)), (45, 670), cv2.FONT_HERSHEY_PLAIN, 10,
(255, 0, 0), 20)
cv2.imshow('OpenPose using OpenCV', frame)
if train_content == '右臂':
right_wrist = point[0]
right_elbow = point[1]
right_shoulder = point[2]
angle_right = Cal_Ang(right_wrist,right_elbow,right_shoulder)
diff = angle_right - angle_last
if diff >=0 and det_1 == 0:
count += 0.5
det_1 = 1
if diff <=0 and det_1 ==1:
count += 0.5
det_1 = 2
cv2.line(frame, tuple(right_wrist), tuple(right_elbow),(0, 0, 255), 3)
cv2.line(frame, tuple(right_elbow), tuple(right_shoulder),(0, 0, 225), 3)
if count % 1.0 == 0.0:
det_1 = 0
print("count:%.2f"%count)
cv2.rectangle(frame, (0, 450), (250, 720), (0, 255, 0), cv2.FILLED)
cv2.putText(frame, str(int(count)), (45, 670), cv2.FONT_HERSHEY_PLAIN, 10,
(255, 0, 0), 20)
cv2.imshow('OpenPose using OpenCV', frame)
cap.release()
【注：这里区分了左臂or右臂】
效果展示【bilibili链接】
