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

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

def read_json(json_path):d = defaultdict(list)for json_name in tqdm(os.listdir(json_path)):json_full_path = json_path + '/' + json_namewith 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

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,vkp = 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

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/2yy = 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/heightproc_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

def split_dataset(image_root,txt_root,img_targe_file,label_target_file,split_ratio=0.6):imgs = os.listdir(image_root)import randomrandom.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 + '/' + imgtxt_path = txt_root + '/' + img.replace('jpg','txt')img_targe_file_full = img_targe_file + '/' + modelabel_target_file_full = label_target_file + '/' + modeshutil.copy(img_path, img_targe_file_full)shutil.copy(txt_path, label_target_file_full)

4.5、验证

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 + '/' + modelabel_path = labels + '/' + modeassert 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

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 iofrom 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,vkp = 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_namewith 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/2yy = 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/heightproc_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 randomrandom.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 + '/' + imgtxt_path = txt_root + '/' + img.replace('jpg','txt')img_targe_file_full = img_targe_file + '/' + modelabel_target_file_full = label_target_file + '/' + modeshutil.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 + '/' + modelabel_path = labels + '/' + modeassert 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！')passif __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