教程 | OpenCV深度神经网络实现人体姿态评估

Posted OpenCV学堂

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了教程 | OpenCV深度神经网络实现人体姿态评估相关的知识,希望对你有一定的参考价值。

OpenCV DNN模块介绍

OpenCV自从发布了DNN模块之后,就开始以开挂的方式支持各种深度学习预训练模型的调用,DNN模块的全称为深度神经网络,但是并不是所有深度学习模型导出到OpenCV DNN模块中都可以使用,只有那些OpenCV声明支持的层与网络模型才会被DNN模块接受,当期OpenCV支持的模型与层类型可以在下面链接中找到相关文档

https://github.com/opencv/opencv/wiki/Deep-Learning-in-OpenCV

模型下载

http://posefs1.perception.cs.cmu.edu/OpenPose/models/pose/coco/pose_iter_440000.caffemodel

http://posefs1.perception.cs.cmu.edu/OpenPose/models/pose/mpi/pose_iter_160000.caffemodel

代码实现

下面只需要如下几步就可以实现基于OpenCV的单人姿态评估:
1.定义COCO数据集支持的18点人体位置与关系位置

BODY_PARTS = { "Nose"0"Neck"1"RShoulder"2"RElbow"3"RWrist"4,
               "LShoulder"5"LElbow"6"LWrist"7"RHip"8"RKnee"9,
               "RAnkle"10"LHip"11"LKnee"12"LAnkle"13"REye"14,
               "LEye"15"REar"16"LEar"17"Background"18 }

POSE_PAIRS = [ ["Neck""RShoulder"], ["Neck""LShoulder"], ["RShoulder""RElbow"],
               ["RElbow""RWrist"], ["LShoulder""LElbow"], ["LElbow""LWrist"],
               ["Neck""RHip"], ["RHip""RKnee"], ["RKnee""RAnkle"], ["Neck""LHip"],
               ["LHip""LKnee"], ["LKnee""LAnkle"], ["Neck""Nose"], ["Nose""REye"],
               ["REye""REar"], ["Nose""LEye"], ["LEye""LEar"] ]

2.定义MPI数据集支持的15点人体位置与关系位置

BODY_PARTS = { "Head": 0, "Neck": 1, "RShoulder": 2, "RElbow": 3, "RWrist": 4,
               "LShoulder": 5, "LElbow": 6, "LWrist": 7, "RHip": 8, "RKnee": 9,
               "RAnkle": 10, "LHip": 11, "LKnee": 12, "LAnkle": 13, "Chest": 14,
               "Background": 15 }

POSE_PAIRS = [ ["Head""Neck"], ["Neck""RShoulder"], ["RShoulder""RElbow"],
               ["RElbow""RWrist"], ["Neck""LShoulder"], ["LShoulder""LElbow"],
               ["LElbow""LWrist"], ["Neck""Chest"], ["Chest""RHip"], ["RHip""RKnee"],
               ["RKnee""RAnkle"], ["Chest""LHip"], ["LHip""LKnee"], ["LKnee""LAnkle"] ]

3.根据不同数据集调用DNN模块加载指定的预训练模型

inWidth = 368
inHeight = 368
thr = 0.1
protoc = "D:/projects/pose_body/mpi/pose_deploy_linevec_faster_4_stages.prototxt"
model = "D:/projects/pose_body/mpi/pose_iter_160000.caffemodel"
net = cv.dnn.readNetFromCaffe(protoc, model)

4.调用OpenCV打开摄像头

cap = cv.VideoCapture(0)
height = cap.get(cv.CAP_PROP_FRAME_HEIGHT)
width = cap.get(cv.CAP_PROP_FRAME_WIDTH)

5.使用前馈网络模型预测

frameWidth = frame.shape[1]
frameHeight = frame.shape[0]
inp = cv.dnn.blobFromImage(frame, 1.0 / 255, (inWidth, inHeight),
                          (0, 0, 0), swapRB=False, crop=False)
net.setInput(inp)
out = net.forward()

6.绘制检测到人体姿态关键点位置

points = []
for i in range(len(BODY_PARTS)):
    # Slice heatmap of corresponging body's part.
    heatMap = out[0, i, :, :]

    # Originally, we try to find all the local maximums. To simplify a sample
    # we just find a global one. However only a single pose at the same time
    # could be detected this way.
    _, conf, _, point = cv.minMaxLoc(heatMap)
    x = (frameWidth * point[0]) / out.shape[3]
    y = (frameHeight * point[1]) / out.shape[2]

    # Add a point if it's confidence is higher than threshold.
    points.append((x, y) if conf > thr else None)

for pair in POSE_PAIRS:
    partFrom = pair[0]
    partTo = pair[1]
    assert(partFrom in BODY_PARTS)
    assert(partTo in BODY_PARTS)

    idFrom = BODY_PARTS[partFrom]
    idTo = BODY_PARTS[partTo]
    if points[idFrom] and points[idTo]:
        x1, y1 = points[idFrom]
        x2, y2 = points[idTo]
        cv.line(frame, (np.int32(x1), np.int32(y1)), (np.int32(x2), np.int32(y2)), (02550), 3)
        cv.ellipse(frame, (np.int32(x1), np.int32(y1)), (33), 00360, (00255), cv.FILLED)
        cv.ellipse(frame, (np.int32(x2), np.int32(y2)), (33), 00360, (00255), cv.FILLED)

完整的代码如下:

import cv2 as cv
import numpy as np


dataset = 'MPI'
if dataset == 'COCO':
    BODY_PARTS = { "Nose"0"Neck"1"RShoulder"2"RElbow"3"RWrist"4,
                   "LShoulder"5"LElbow"6"LWrist"7"RHip"8"RKnee"9,
                   "RAnkle"10"LHip"11"LKnee"12"LAnkle"13"REye"14,
                   "LEye"15"REar"16"LEar"17"Background"18 }

    POSE_PAIRS = [ ["Neck""RShoulder"], ["Neck""LShoulder"], ["RShoulder""RElbow"],
                   ["RElbow""RWrist"], ["LShoulder""LElbow"], ["LElbow""LWrist"],
                   ["Neck""RHip"], ["RHip""RKnee"], ["RKnee""RAnkle"], ["Neck""LHip"],
                   ["LHip""LKnee"], ["LKnee""LAnkle"], ["Neck""Nose"], ["Nose""REye"],
                   ["REye""REar"], ["Nose""LEye"], ["LEye""LEar"] ]
else:
    assert(dataset == 'MPI')
    BODY_PARTS = { "Head"0"Neck"1"RShoulder"2"RElbow"3"RWrist"4,
                   "LShoulder"5"LElbow"6"LWrist"7"RHip"8"RKnee"9,
                   "RAnkle"10"LHip"11"LKnee"12"LAnkle"13"Chest"14,
                   "Background"15 }

    POSE_PAIRS = [ ["Head""Neck"], ["Neck""RShoulder"], ["RShoulder""RElbow"],
                   ["RElbow""RWrist"], ["Neck""LShoulder"], ["LShoulder""LElbow"],
                   ["LElbow""LWrist"], ["Neck""Chest"], ["Chest""RHip"], ["RHip""RKnee"],
                   ["RKnee""RAnkle"], ["Chest""LHip"], ["LHip""LKnee"], ["LKnee""LAnkle"] ]

inWidth = 368
inHeight = 368
thr = 0.1
protoc = "D:/projects/pose_body/mpi/pose_deploy_linevec_faster_4_stages.prototxt"
model = "D:/projects/pose_body/mpi/pose_iter_160000.caffemodel"
net = cv.dnn.readNetFromCaffe(protoc, model)

cap = cv.VideoCapture(0)
height = cap.get(cv.CAP_PROP_FRAME_HEIGHT)
width = cap.get(cv.CAP_PROP_FRAME_WIDTH)
video_writer = cv.VideoWriter("D:/pose_estimation_demo.mp4", cv.VideoWriter_fourcc('D''I''V''X'), 15, (640480), True)
while cv.waitKey(1) < 0:
    hasFrame, frame = cap.read()
    if not hasFrame:
        cv.waitKey()
        break

    frameWidth = frame.shape[1]
    frameHeight = frame.shape[0]
    inp = cv.dnn.blobFromImage(frame, 1.0 / 255, (inWidth, inHeight),
                              (000), swapRB=False, crop=False)
    net.setInput(inp)
    out = net.forward()

    print(len(BODY_PARTS), out.shape[0])
    # assert(len(BODY_PARTS) == out.shape[1])

    points = []
    for i in range(len(BODY_PARTS)):
        # Slice heatmap of corresponging body's part.
        heatMap = out[0, i, :, :]

        # Originally, we try to find all the local maximums. To simplify a sample
        # we just find a global one. However only a single pose at the same time
        # could be detected this way.
        _, conf, _, point = cv.minMaxLoc(heatMap)
        x = (frameWidth * point[0]) / out.shape[3]
        y = (frameHeight * point[1]) / out.shape[2]

        # Add a point if it's confidence is higher than threshold.
        points.append((x, y) if conf > thr else None)

    for pair in POSE_PAIRS:
        partFrom = pair[0]
        partTo = pair[1]
        assert(partFrom in BODY_PARTS)
        assert(partTo in BODY_PARTS)

        idFrom = BODY_PARTS[partFrom]
        idTo = BODY_PARTS[partTo]
        if points[idFrom] and points[idTo]:
            x1, y1 = points[idFrom]
            x2, y2 = points[idTo]
            cv.line(frame, (np.int32(x1), np.int32(y1)), (np.int32(x2), np.int32(y2)), (02550), 3)
            cv.ellipse(frame, (np.int32(x1), np.int32(y1)), (33), 00360, (00255), cv.FILLED)
            cv.ellipse(frame, (np.int32(x2), np.int32(y2)), (33), 00360, (00255), cv.FILLED)

    t, _ = net.getPerfProfile()
    freq = cv.getTickFrequency() / 1000
    cv.putText(frame, '%.2fms' % (t / freq), (1020), cv.FONT_HERSHEY_SIMPLEX, 0.5, (000))
    # video_writer.write(frame);
    # cv.imwrite("D:/pose.png", frame)
    cv.imshow('OpenPose using OpenCV', frame)

运行结果如下:

【推荐阅读】











坚其志,苦其心,劳其力

事无大小,必有所成


关注【OpenCV学堂】

长按或者扫码即可关注

以上是关于教程 | OpenCV深度神经网络实现人体姿态评估的主要内容,如果未能解决你的问题,请参考以下文章

目标跟踪的深度学习方法与opencv实现

OpenPose 安装和使用教程

「深度学习一遍过」必修27:基于Mask-RCNN的人体姿态估计的设计与实现

「深度学习一遍过」必修27:基于Mask-RCNN的人体姿态估计的设计与实现

「深度学习一遍过」必修27:基于Mask-RCNN的人体姿态估计的设计与实现

MidiaPipe +stgcn(时空图卷积网络)实现人体姿态判断(单目标)