【数据集】Yolo人体关键点数据集处理

鲸可落2024-08-09 8:20

文章目录

1、介绍

人体关键点检测（Human Keypoints Detection）又称为人体姿态估计2D Pose，是计算机视觉中一个相对基础的任务，是人体动作识别、行为分析、人机交互等的前置任务。一般情况下可以将人体关键点检测细分为单人/多人关键点检测、2D/3D关键点检测，同时有算法在完成关键点检测之后还会进行关键点的跟踪，也被称为人体姿态跟踪。

本次要介绍的数据集是2D关键点检测数据集，数据集主要来自COCO2017，经过对COCO数据集JSON文件进行预处理提取人体关键点信息，一共提取10000 张人体姿态数据集，以及对应的必要信息，已经转化为Yolo格式存储。

2、数据集格式

复制代码 
keypoint_dataset
	|
	|_______images
	|		  |_____000000000049.jpg
	|		  |_____ ......
	|_______labels
	|		  |_____000000000049.txt
	|		  |_____ ......

3 、COCO人体关键点示意图

下图中，共17个关节点（鼻子x1、眼睛x2、耳朵x2、肩部x2、肘部x2、手腕x2、髋部x2、膝关节x2、脚腕x2）：

4、数据集预处理

这里数据集预处理主要包括5个内容，分别是read_json、show、json2yolo、split_dataset、validation。

4.1、读取JSON文件

python 复制代码 
def read_json(json_path):
    d = defaultdict(list)
    for json_name in tqdm(os.listdir(json_path)):
        json_full_path = json_path + '/' + json_name
        with open(json_full_path, 'r', encoding='utf-8') as file:  
            data = json.load(file)
        image_name = data['image_name']
        # category = data['category']
        # width = data['width']
        # height = data['height']
        # keypoints = data['key_points']
        # bbox = data['bbox']
        image_id = image_name.split('.')[0]
        d[int(image_id)]= data 
    print(f'json 文件读取完成,一共{len(d.keys())}个数据')
    return d

4.2、可视化

python 复制代码 
def visualization(ax,keypoints,bbox):
    #随机颜色
    c = (np.random.random((1, 3))*0.6+0.4).tolist()[0]
    #关键点之间的连线
    ls = [[15,13],[13,11],[16,14],[14,12],[11,12],[5,11],[6,12],[5,6],[5,7],
          [6,8],[7,9],[8,10],[1,2],[0,1],[0,2],[1,3],[2,4],[3,5],[4,6]]
    sks = np.array(ls)
    #获取关键点坐标x,y,v
    kp = np.array(keypoints)
    x = kp[0::3]
    y = kp[1::3]
    v = kp[2::3]
    for sk in sks:
        if np.all(v[sk]>0):
            # 画点之间的连接线
            plt.plot(x[sk],y[sk], linewidth=1, color=c)
    # 画点
    p = plt.plot(x[v>0], y[v>0],'o',markersize=4, markerfacecolor=c, markeredgecolor='k',markeredgewidth=1)
    p = plt.plot(x[v>1], y[v>1],'o',markersize=4, markerfacecolor=c, markeredgecolor=c, markeredgewidth=1)
    # 画矩形边界,多边形填充+矩形边界:
    x, y, w, h = bbox[0],bbox[1],bbox[2],bbox[3]
    ax.add_patch(Polygon(xy=[[x, y], [x, y+h], [x+w, y+h], [x+w, y]], color='k', alpha=0.3))
    ax.add_patch(Rectangle(xy=(x, y), width=w, height=h, fill=False, color=c, alpha=1))
    plt.plot(x + w/2,y+h/2,'*',markersize=5, markerfacecolor=c, markeredgecolor=c, markeredgewidth=1)

def show(data,image_root):
    for image_id in data.keys():

        img = io.imread('%s/%s' % (image_root, data[image_id]['image_name']))
        plt.axis('off')
        ax = plt.gca()
        for i in range(len(data[image_id]['key_points'])):
            visualization(ax, data[image_id]['key_points'][i], data[image_id]['bbox'][i])
        plt.imshow(img)
        plt.axis('off')
        plt.show()

4.3、JSON格式转Yolo格式

yolo标注格式为：类别 、标注框中心点(x,y)、长和宽(w,h)，关键点坐标以及可见度(kx1,ky1,kv1,kx2,ky2,kv2...)，然后并根据图片长宽进行归一化处理

python 复制代码 
def json2yolo(data,txt_path):
    if not os.path.exists(txt_path):
        os.mkdir(txt_path)
    for image_id in tqdm(data.keys()):
        txt_name = data[image_id]['image_name'].split('.')[0]
        category = data[image_id]['category']
        width = data[image_id]['width']
        height = data[image_id]['height']
        keypoints = data[image_id]['key_points']
        bbox = data[image_id]['bbox']
        '''
        yolo标注格式
        类别 、标注框中心点、长和宽，关键点坐标以及可见度
        根据图片长高进行归一化处理
        '''
        for i in range(len(keypoints)):
            #对标注框进行预处理
            box = bbox[i]
            keypoint = keypoints[i]
            # print(keypoint)
            #x,y为左上角坐标，w,h为宽高
            x,y,w,h = box[0],box[1],box[2],box[3]
            xx = x + w/2
            yy = y + h/2
            #归一化
            xx,yy,ww,hh = xx/width,yy/height,w/width,h/height
            #关键点归一化
            proc_keypoint = []
            for p in range(0,len(keypoint),3):
                x,y,v = keypoint[p],keypoint[p+1],keypoint[p+2]
                kx,ky = x/width, y/height
                proc_keypoint.extend([kx,ky,v])
            #写入txt文件
            yolo_str = f'{category} {xx:.6f} {yy:.6f} {ww:.6f} {hh:.6f} '
            yolo_str = yolo_str + ' '.join([f'{i:.6f}' for i in proc_keypoint])
            
            with open(f'{txt_path}/{txt_name}.txt','a+') as f:
                f.write(yolo_str + '\n')
    print('转化为yolo txt格式完成!')

4.4、划分数据集

划分数据集格式为yolo需要的格式

python 复制代码 
def split_dataset(image_root,txt_root,img_targe_file,label_target_file,split_ratio=0.6):
    imgs = os.listdir(image_root)
    import random
    random.seed(2024)
    random.shuffle(imgs)
    #这里仅仅取了1万张图片进行测试，由于显存限制，
    ims = imgs[:10000]
    random.shuffle(ims)
    train_num = int(len(ims)*split_ratio)
    val_num = int(len(ims)*(1-split_ratio)/2)

    train_set = ims[:train_num]
    val_set = ims[train_num:train_num+val_num]
    test_set = ims[-val_num:]

    move_file(train_set,image_root,txt_root,img_targe_file,label_target_file,mode='train')
    move_file(val_set,image_root,txt_root,img_targe_file,label_target_file,mode='val')
    move_file(test_set,image_root,txt_root,img_targe_file,label_target_file,mode='test')

def move_file(dataset,image_root,txt_root,img_targe_file,label_target_file,mode='train'):
    for img in tqdm(dataset):
        img_path = image_root + '/' + img
        txt_path = txt_root + '/' + img.replace('jpg','txt')
        img_targe_file_full = img_targe_file + '/' + mode
        label_target_file_full = label_target_file + '/' + mode


        shutil.copy(img_path, img_targe_file_full)
        shutil.copy(txt_path, label_target_file_full)

4.5、验证

python 复制代码 
def validation():
    images = r'E:\datasets\keypoint_dataset\datasets\images'
    labels = r'E:\datasets\keypoint_dataset\datasets\labels'
    for mode in ['train','val','test']:
        img_path = images + '/' + mode
        label_path = labels + '/' + mode
        assert len(os.listdir(img_path)) == len(os.listdir(label_path)),f'{mode}数据划分验证失败'
        for img,lb in zip(os.listdir(img_path),os.listdir(label_path)):
            img_name = img.split('.')[0]
            lb_name = lb.split('.')[0]
            if img_name != lb_name:
                 assert '数据划分验证失败'
    print('数据划分验证成功！')
    pass

5、完整代码

python 复制代码 
import json
from collections import defaultdict
import numpy as np
import os
from tqdm import tqdm
import shutil
from matplotlib import pyplot as plt
import skimage.io as io

from matplotlib.patches import Polygon,Rectangle
def visualization(ax,keypoints,bbox):
    #随机颜色
    c = (np.random.random((1, 3))*0.6+0.4).tolist()[0]
    #关键点之间的连线
    ls = [[15,13],[13,11],[16,14],[14,12],[11,12],[5,11],[6,12],[5,6],[5,7],
          [6,8],[7,9],[8,10],[1,2],[0,1],[0,2],[1,3],[2,4],[3,5],[4,6]]
    sks = np.array(ls)
    #获取关键点坐标x,y,v
    kp = np.array(keypoints)
    x = kp[0::3]
    y = kp[1::3]
    v = kp[2::3]
    for sk in sks:
        if np.all(v[sk]>0):
            # 画点之间的连接线
            plt.plot(x[sk],y[sk], linewidth=1, color=c)
    # 画点
    p = plt.plot(x[v>0], y[v>0],'o',markersize=4, markerfacecolor=c, markeredgecolor='k',markeredgewidth=1)
    p = plt.plot(x[v>1], y[v>1],'o',markersize=4, markerfacecolor=c, markeredgecolor=c, markeredgewidth=1)
    # 画矩形边界,多边形填充+矩形边界:
    x, y, w, h = bbox[0],bbox[1],bbox[2],bbox[3]
    ax.add_patch(Polygon(xy=[[x, y], [x, y+h], [x+w, y+h], [x+w, y]], color='k', alpha=0.3))
    ax.add_patch(Rectangle(xy=(x, y), width=w, height=h, fill=False, color=c, alpha=1))
    plt.plot(x + w/2,y+h/2,'*',markersize=5, markerfacecolor=c, markeredgecolor=c, markeredgewidth=1)

def read_json(json_path):
    d = defaultdict(list)
    
    for json_name in tqdm(os.listdir(json_path)):
        json_full_path = json_path + '/' + json_name
        with open(json_full_path, 'r', encoding='utf-8') as file:  
            data = json.load(file)
        image_name = data['image_name']
        # category = data['category']
        # width = data['width']
        # height = data['height']
        # keypoints = data['key_points']
        # bbox = data['bbox']
        image_id = image_name.split('.')[0]
        
        d[int(image_id)]= data 
    
    print(f'json 文件读取完成,一共{len(d.keys())}个数据')
    return d
def show(data,image_root):
    for image_id in data.keys():

        img = io.imread('%s/%s' % (image_root, data[image_id]['image_name']))
        plt.axis('off')
        ax = plt.gca()
        for i in range(len(data[image_id]['key_points'])):
            visualization(ax, data[image_id]['key_points'][i], data[image_id]['bbox'][i])
        plt.imshow(img)
        plt.axis('off')
        plt.show()
def json2yolo(data,txt_path):
    if not os.path.exists(txt_path):
        os.mkdir(txt_path)
    for image_id in tqdm(data.keys()):
        txt_name = data[image_id]['image_name'].split('.')[0]
        category = data[image_id]['category']
        width = data[image_id]['width']
        height = data[image_id]['height']
        keypoints = data[image_id]['key_points']
        bbox = data[image_id]['bbox']
        '''
        yolo标注格式
        类别 、标注框中心点、长和宽，关键点坐标以及可见度
        根据图片长高进行归一化处理
        '''
        for i in range(len(keypoints)):
            #对标注框进行预处理
            box = bbox[i]
            keypoint = keypoints[i]
            # print(keypoint)
            #x,y为左上角坐标，w,h为宽高
            x,y,w,h = box[0],box[1],box[2],box[3]
            xx = x + w/2
            yy = y + h/2
            #归一化
            xx,yy,ww,hh = xx/width,yy/height,w/width,h/height
            #关键点归一化
            proc_keypoint = []
            for p in range(0,len(keypoint),3):
                x,y,v = keypoint[p],keypoint[p+1],keypoint[p+2]
                kx,ky = x/width, y/height
                proc_keypoint.extend([kx,ky,v])
            #写入txt文件
            yolo_str = f'{category} {xx:.6f} {yy:.6f} {ww:.6f} {hh:.6f} '
            yolo_str = yolo_str + ' '.join([f'{i:.6f}' for i in proc_keypoint])
            
            with open(f'{txt_path}/{txt_name}.txt','a+') as f:
                f.write(yolo_str + '\n')
    print('转化为yolo txt格式完成!')

def split_dataset(image_root,txt_root,img_targe_file,label_target_file,split_ratio=0.6):
    imgs = os.listdir(image_root)
    import random
    random.seed(2024)
    random.shuffle(imgs)
    #这里仅仅取了1万张图片进行测试，由于显存限制，
    ims = imgs[:10000]
    random.shuffle(ims)
    train_num = int(len(ims)*split_ratio)
    val_num = int(len(ims)*(1-split_ratio)/2)

    train_set = ims[:train_num]
    val_set = ims[train_num:train_num+val_num]
    test_set = ims[-val_num:]

    move_file(train_set,image_root,txt_root,img_targe_file,label_target_file,mode='train')
    move_file(val_set,image_root,txt_root,img_targe_file,label_target_file,mode='val')
    move_file(test_set,image_root,txt_root,img_targe_file,label_target_file,mode='test')

def move_file(dataset,image_root,txt_root,img_targe_file,label_target_file,mode='train'):
    for img in tqdm(dataset):
        img_path = image_root + '/' + img
        txt_path = txt_root + '/' + img.replace('jpg','txt')
        img_targe_file_full = img_targe_file + '/' + mode
        label_target_file_full = label_target_file + '/' + mode


        shutil.copy(img_path, img_targe_file_full)
        shutil.copy(txt_path, label_target_file_full)
def validation():
    images = r'E:\datasets\keypoint_dataset\datasets\images'
    labels = r'E:\datasets\keypoint_dataset\datasets\labels'
    for mode in ['train','val','test']:
        img_path = images + '/' + mode
        label_path = labels + '/' + mode
        assert len(os.listdir(img_path)) == len(os.listdir(label_path)),f'{mode}数据划分验证失败'
        for img,lb in zip(os.listdir(img_path),os.listdir(label_path)):
            img_name = img.split('.')[0]
            lb_name = lb.split('.')[0]
            if img_name != lb_name:
                 assert '数据划分验证失败'
    print('数据划分验证成功！')
    pass

if __name__ == '__main__':
    root = 'E:/datasets/keypoint_dataset'
    image_root = 'E:/datasets/keypoint_dataset/images'
    json_root = 'E:/datasets/keypoint_dataset/labels'
    txt_root = root + '/'+ 'txt'
    #读取json文件
    data = read_json(json_root)
    #可视化
    # show(data,image_root)
    #json2yolo
    # json2yolo(data,txt_root)
    #split dataset
    # img_target_file = root + '/'+ 'datasets' + '/images'
    # label_target_file = root + '/'+ 'datasets' + '/labels'
    # split_dataset(image_root,txt_root,img_target_file,label_target_file,split_ratio=0.7)
    #validation验证
    validation()

6、数据集下载链接

数据集下载链接

相关推荐
羊羊小栈2 天前
基于「YOLO目标检测 + 多模态AI分析」的遥感影像目标检测分析系统(vue+flask+数据集+模型训练)
人工智能·深度学习·yolo·目标检测·毕业设计·大作业
大霸王龙3 天前
基于vLLM与YOLO的智能图像分类系统
yolo·分类·数据挖掘
m_136874 天前
Mac M 系列芯片 YOLOv8 部署教程（CPU/Metal 后端一键安装）
yolo·macos
格林威4 天前
机器视觉在半导体制造中有哪些检测应用
人工智能·数码相机·yolo·计算机视觉·视觉检测·制造·相机
、、、、南山小雨、、、、5 天前
YOLO在ubuntu22安装
yolo
羊羊小栈5 天前
基于「YOLO目标检测 + 多模态AI分析」的铁路轨道缺陷检测安全系统(vue+flask+数据集+模型训练)
人工智能·yolo·目标检测·语言模型·毕业设计·创业创新·大作业
Python图像识别5 天前
63_基于深度学习的草莓病害检测识别系统（yolo11、yolov8、yolov5+UI界面+Python项目源码+模型+标注好的数据集）
python·深度学习·yolo
范男5 天前
YOLO11目标检测运行推理简约GUI界面
图像处理·人工智能·yolo·计算机视觉·视觉检测
model20055 天前
ubuntu24.04＋5070ti训练yolo模型(2)
人工智能·yolo
强盛小灵通专卖员6 天前
RK3576边缘计算设备部署YOLOv11
人工智能·深度学习·yolo·边缘计算·sci·rk3576·小论文
热门推荐
01UV 工具安装与国内镜像源配置指南02GitHub 镜像站点03Claude Code 平替：OpenAI发布 Codex CLI ，GPT-5 国内直接使用0446个Nano-banana 精选提示词，持续更新中05UV安装并设置国内源06保姆级教程：手把手教你用Dify实现完美多轮对话（附Chatflow和提示词）07解决 WSL Ubuntu 中 /etc/resolv.conf 自动重置问题08A股预测还能更准？开源大模型Kronos带你跑通预测+回测全流程09Spec-Kit 使用指南10KGG转MP3工具|非KGM文件|解密音频