[PaddleHub]全民健身热潮之AI帮你仰卧起坐计数

转自AI Studio，原文链接：[畊宏女孩]全民健身热潮之AI帮你仰卧起坐计数 - 飞桨AI Studio1.背景介绍自从疫情以来，全民健身横行~ 大家居家锻炼，无疑需要用到卷腹、仰卧起坐等等这些室内健身方法。为了方便做的时候不要再操心计数的问题，从而诞生了本产品，AI帮你仰卧起坐计数2.实现思路1.用户打开手机APP，将手机固定在场地一侧，适当设置手机角度，根据应用的自动语音提示调整身体与手

AI Studio

1605人浏览 · 2022-05-12 22:17:52

AI Studio · 2022-05-12 22:17:52 发布

转自AI Studio，原文链接：

[畊宏女孩]全民健身热潮之AI帮你仰卧起坐计数 - 飞桨 AI Studio

1.背景介绍

自从疫情以来，全民健身横行~ 大家居家锻炼，无疑需要用到卷腹、仰卧起坐等等这些室内健身方法。为了方便做的时候不要再操心计数的问题，从而诞生了本产品，AI帮你仰卧起坐计数

2.实现思路

1.用户打开手机APP，将手机固定在场地一侧，适当设置手机角度，根据应用的自动语音提示调整身体与手机距离，直到人体完全位于识别框内，即可开始运动。
2.通过PaddleHub的human_pose_estimation_resnet50_mpii模型，进行人体关键点检测。
3.根据检测的数据计数（此处选择头部关键点进行判断，一次完整的上下为一次仰卧）

二、环境准备

1.PaddleHub安装

In [1]

!pip install -U pip --user >log.log
!pip install -U paddlehub >log.log

WARNING: You are using pip version 22.0.4; however, version 22.1 is available.
You should consider upgrading via the '/opt/conda/envs/python35-paddle120-env/bin/python -m pip install --upgrade pip' command.

In [ ]

!pip list |grep paddle

paddle2onnx            0.9.5
paddlehub              2.2.0
paddlenlp              2.0.7
paddlepaddle-gpu       2.2.2.post101
tb-paddle              0.3.6

2.human_pose_estimation_resnet50_mpii模型安装

模型地址：飞桨PaddlePaddle-源于产业实践的开源深度学习平台
模型概述：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。

In [2]

!hub install human_pose_estimation_resnet50_mpii >log.log

[2022-05-12 15:07:19,455] [    INFO] - Successfully installed human_pose_estimation_resnet50_mpii-1.1.1

In [ ]

|human_pose_estimation_res|/home/aistudio/.paddlehub/modules/human_pose_estim|

三、人体关键点检测示例

1.演示视频

自行获取视频

2.关键点检测演示

针对下面这三张图片做关键点检测，具体如下：

In [ ]

import cv2
import paddlehub as hub

pose_estimation = hub.Module(name="human_pose_estimation_resnet50_mpii")

image1=cv2.imread('work/1.png') # 坐直
image2=cv2.imread('work/2.png') # 全躺
image3=cv2.imread('work/3.png') # 中间状态
results = pose_estimation.keypoint_detection(images=[image1,image2,image3], visualization=True)

[2022-05-09 16:54:52,291] [ WARNING] - The _initialize method in HubModule will soon be deprecated, you can use the __init__() to handle the initialization of the object
W0509 16:54:52.295652   101 analysis_predictor.cc:1350] Deprecated. Please use CreatePredictor instead.

image saved in output_pose/ndarray_time=1652086492754842.jpg
image saved in output_pose/ndarray_time=1652086492754871.jpg
image saved in output_pose/ndarray_time=1652086492754875.jpg

In [ ]

# 打印关键点 
print(results[0]['data'])

OrderedDict([('left_ankle', [534, 305]), ('left_knee', [461, 213]), ('left_hip', [380, 289]), ('right_hip', [388, 269]), ('right_knee', [461, 209]), ('right_ankle', [527, 301]), ('pelvis', [388, 277]), ('thorax', [351, 176]), ('upper_neck', [366, 144]), ('head_top', [388, 80]), ('right_wrist', [402, 144]), ('right_elbow', [446, 209]), ('right_shoulder', [351, 184]), ('left_shoulder', [366, 168]), ('left_elbow', [439, 205]), ('left_wrist', [410, 144])])

查看output_pose 下输出的图片：

3.如何判断起身下落

判断一次仰卧起坐的依据是什么呢？根据上面的三张图，可以轻松得出结论，用头部坐标的移动可以作为评判标准。

In [ ]

# 打印三张头部关键点 
print(results[0]['data']['head_top'])
print(results[1]['data']['head_top'])
print(results[2]['data']['head_top'])

[388, 80]
[87, 305]
[301, 107]

四、智能计数

改进：采用GPU预测

10秒的视频大概：运行时长: 18秒

70秒的视频大概：运行时长: 70秒

速度起飞~

In [4]

import cv2
import paddlehub as hub
import math
from matplotlib import pyplot as plt
import numpy as np
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
%matplotlib inline

def countYwqz():
    pose_estimation = hub.Module(name="human_pose_estimation_resnet50_mpii")

    flag = False
    count = 0
    num = 0
    all_num = []
    flip_list = []
    fps = 60
    # 可选择web视频流或者文件
    file_name = 'work/ywqz.mp4'
    cap = cv2.VideoCapture(file_name)
    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    # out后期可以合成视频返回
    out = cv2.VideoWriter(
        'output.mp4',
        fourcc,
        fps,
        (width,height))

    while cap.isOpened():
        success, image = cap.read()
        # print(image)
        if not success:
            break
        image_height, image_width, _ = image.shape
        # print(image_height, image_width)

        image.flags.writeable = False
        results = pose_estimation.keypoint_detection(images=[image], visualization=True, use_gpu=True)

        flip = results[0]['data']['head_top'][1] # 获取头部的y轴坐标值
        flip_list.append(flip)
        all_num.append(num)
        num +=1
    
    # 写入视频
    img_root="output_pose/"
    # 排序，不然是乱序的合成出来
    im_names=os.listdir(img_root)  
    im_names.sort(key=lambda x: int(x.replace("ndarray_time=","").split('.')[0]))
    for im_name in range(len(im_names)):
        img = img_root+str(im_names[im_name])
        print(img)
        frame=cv2.imread(img)
        out.write(frame)  
    out.release()

    return all_num,flip_list

def get_count(x,y):
    count = 0
    flag = False
    count_list = [0] # 记录极值的y值
    for i in range(len(y)-1):
        if y[i] <= y[i + 1] and flag == False:
            continue
        elif y[i] >= y[i + 1] and flag == True:
            continue
        else:
            # 防止附近的轻微抖动也被计入数据
            if abs(count_list[-1] - y[i]) >100 or abs(count_list[-1] - y[i-1]) >100 or abs(count_list[-1] - y[i-2]) >100 or abs(count_list[-1] - y[i-3]) >100 or abs(count_list[-1] - y[i+1]) >100  or abs(count_list[-1] - y[i+2]) >100  or abs(count_list[-1] - y[i+3]) >100:
                count = count + 1
                count_list.append(y[i])
                print(x[i])
                flag = not flag
    return math.floor(count/2)
    

if __name__ == "__main__":
    x,y = countYwqz()

    plt.figure(figsize=(8, 8))
    count = get_count(x,y)
    plt.title(f"point numbers: {count}")
    plt.plot(x, y)
    plt.show()