在复现st-gcn时，需要用到Openpose提取视频中的骨骼信息，但是使用的时候我却遇到了很多问题。
准备工作
下载、编译openpose，这个网上也有很多教程，这里我不赘述，参考openpose安装到实战（win10）
调用Openpsoe
import os
import shutil
from pathlib import Path
import json
from .utils import get_frames
def extract_skeleon(path_data, path_out, label, index,
openposeCmd='D:/githouse/openpose/build/x64/Release/OpenPoseDemo.exe',
model='COCO'):
'''
利用OpenPoseDemo.exe处理指定路径下的视频文件，生成.json的骨骼文件以及处理后的视频
param
path_data: 视频路径
path_out: 保存的结果的路径
label: 视频标签
index: 标签索引
return
'''
videos = os.listdir(path_data)
path_out_snippets = path_out + '/snippets'
if not os.path.exists(path_out_snippets):
os.makedirs(path_out_snippets)
path_out_sequence = path_out + '/data'
if not os.path.exists(path_out_sequence):
os.makedirs(path_out_sequence)
for video in videos:
path_video = '{}/{}'.format(path_data, video)
video_name = video.split('.')[0]
## 姿态估计，输出处理后的视频
path_out_video = '{}/{}'.format(path_out_snippets, video_name)
openpose_args = dict(
video=path_video,
write_json=path_out_video,
display=0,
render_pose=0,
model_pose=model)
command_line = openposeCmd + ' '
command_line += ' '.join(['--{} {}'.format(k, v) for k, v in openpose_args.items()])
shutil.rmtree(path_out_video, ignore_errors=True)
print('执行指令:', command_line)
os.system(command_line)
## 生成骨骼数据文件
video = get_frames(path_video)
height, width, _ = video[0].shape
# 这里可以修改label, label_index
video_info = json_pack(path_out_video, video_name, width, height, label, index)
path_out_json = path_out_sequence+'/{}.json'.format(video_name)
with open(path_out_json, 'w') as outfile:
json.dump(video_info, outfile)
if len(video_info['data']) == 0:
print('{} Can not find pose estimation results.'.format(video))
else:
print('{} pose estimation complete.'.format(video))
return
def json_pack(snippets_dir, video_name, frame_width, frame_height, label='unknown', label_index=-1):
'''
将 视频提取的 *.json数据打包成dict
param
snippets_dir:
video_name:
frame_width:
frame_height:
label:
label_index:
return:
video_info:dict{
data:list(dict{
skeleton:list
pose:2d坐标[x,y]
score:坐标的置信分数
})
label:
label_index:
}
'''
sequence_info = []
p = Path(snippets_dir)
for path in p.glob(video_name + '*.json'):
json_path = str(path)
print(path)
frame_id = int(path.stem.split('_')[-2])
frame_data = {'frame_index': frame_id}
data = json.load(open(json_path))
skeletons = []
for person in data['people']:
score, coordinates = [], []
skeleton = {}
keypoints = person['pose_keypoints_2d']
for i in range(0, len(keypoints), 3):
coordinates += [keypoints[i]/frame_width, keypoints[i + 1]/frame_height]
score += [keypoints[i + 2]]
skeleton['pose'] = coordinates
skeleton['score'] = score
skeletons += [skeleton]
frame_data['skeleton'] = skeletons
sequence_info += [frame_data]
video_info = dict()
video_info['data'] = sequence_info
video_info['label'] = label
video_info['label_index'] = label_index
return video_info问题
问题1：路径有空格
但是我遇到了一个问题，输出提示我
Can not find pose estimation results
我将指令复制，打开cmd执行，同样报错，指令如下：
D:/githouse/openpose/build/x64/Release/OpenPoseDemo.exe --video E:/DataSet/test data/resize/boxing/person01_boxing_d1_uncomp(003320-021320).avi --write_json E:/DataSet/test data/skeleton/boxing/snippets/person01_boxing_d1_uncomp(003320-021320) --display 0 --render_pose 0 --model_pose COCO
根据命令行的提示我找到了原因，路径中有空格，导致路径无法被识别
问题2：找不到*.dll
我将数据集路径以及代码修改后，再次运行还是出现了
Can not find pose estimation results
同样我在cmd中执行，这次提示找不到glob.dll,...，反正就是找不到很多.dll文件，经过一番操作，我发现这些文件在D:/githouse/openpose/build/bin中找到了，于是我把这些文件复制到OpenposeDemo.exe所在的目录，我这里是：
D:/githouse/openpose/build/x64/Release/OpenPoseDemo.exe
然后该问题就没出现了
问题3：工作路径没有设置正确
解决了前两个问题，运行还是报错，这次命令行给我三个可能的原因：
（1）openpose需要的包没有下载
（2）工作路径没有在model文件夹所在路径
（3）文件路径有空格
一眼看1,3是不可能的，只能是2了，于是我在pycharm中把工作路径设置为model文件夹的父路径，即openpose的安装路径。
问题解决了，运行成功，nice。这里附上全部代码
utils.py
import os
import skvideo
# skvideo.setFFmpegPath('S:/ffmpeg/bin')
import skvideo.io
import numpy as np
import cv2
def parse_info(video_info, num_person_in=5, num_person_out=2):
data_numpy = np.zeros((3, len(video_info['data']), 18, num_person_in))
for frame_info in video_info['data']:
frame_index = frame_info['frame_index']
for m, skeleton_info in enumerate(frame_info["skeleton"]):
if m >= num_person_in:
break
pose = skeleton_info['pose']
score = skeleton_info['score']
data_numpy[0, frame_index, :, m] = pose[0::2]
data_numpy[1, frame_index, :, m] = pose[1::2]
data_numpy[2, frame_index, :, m] = score
# centralization
data_numpy[0:2] = data_numpy[0:2] - 0.5
data_numpy[0][data_numpy[2] == 0] = 0
data_numpy[1][data_numpy[2] == 0] = 0
sort_index = (-data_numpy[2, :, :, :].sum(axis=1)).argsort(axis=1)
for t, s in enumerate(sort_index):
data_numpy[:, t, :, :] = data_numpy[:, t, :, s].transpose((1, 2,
0))
data_numpy = data_numpy[:, :, :, :num_person_out]
label = video_info['label_index']
return data_numpy, label
def get_frames(video_path):
vread = skvideo.io.vread(video_path)
video = []
for frame in vread:
video.append(frame)
return video
def play(video_path, fps=30):
cap = cv2.VideoCapture(video_path)
while(cap.isOpened()):
ret, frame = cap.read()
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
cv2.imshow('frame', gray)
if cv2.waitKey(1000/fps) & 0xFF == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
def resize(path_data, path_out, size = (340,256)):
'''
将输入路径下的视频大小调整为指定大小，默认为340x256，并将修改后的视频保存到指定路径
param
path_data:
path_out:
'''
videos = os.listdir(path_data)
size_str = '{}x{}'.format(size[0], size[1])
for video in videos:
video_path = '{}/{}'.format(path_data, video)
out_path = '{}/{}'.format(path_out, video)
writer = skvideo.io.FFmpegWriter(out_path,
outputdict={'-f': 'mp4', '-vcodec': 'libx264', '-s': size_str, '-r': '30'})
reader = skvideo.io.FFmpegReader(video_path)
for frame in reader.nextFrame():
writer.writeFrame(frame)
writer.close()
print('{} resize success'.format(video))skeleton.py
import os
import shutil
from pathlib import Path
import json
from .utils import get_frames
def extract_skeleon(path_data, path_out, label, index,
openposeCmd='D:/githouse/openpose/build/x64/Release/OpenPoseDemo.exe',
model='COCO'):
'''
利用OpenPoseDemo.exe处理指定路径下的视频文件，生成.json的骨骼文件以及处理后的视频
param
path_data: 视频路径
path_out: 保存的结果的路径
label: 视频标签
index: 标签索引
return
'''
videos = os.listdir(path_data)
path_out_snippets = path_out + '/snippets'
if not os.path.exists(path_out_snippets):
os.makedirs(path_out_snippets)
path_out_sequence = path_out + '/data'
if not os.path.exists(path_out_sequence):
os.makedirs(path_out_sequence)
for video in videos:
path_video = '{}/{}'.format(path_data, video)
video_name = video.split('.')[0]
## 姿态估计，输出处理后的视频
path_out_video = '{}/{}'.format(path_out_snippets, video_name)
openpose_args = dict(
video=path_video,
write_json=path_out_video,
display=0,
render_pose=0,
model_pose=model)
command_line = openposeCmd + ' '
command_line += ' '.join(['--{} {}'.format(k, v) for k, v in openpose_args.items()])
shutil.rmtree(path_out_video, ignore_errors=True)
print('执行指令:', command_line)
os.system(command_line)
## 生成骨骼数据文件
video = get_frames(path_video)
height, width, _ = video[0].shape
# 这里可以修改label, label_index
video_info = json_pack(path_out_video, video_name, width, height, label, index)
path_out_json = path_out_sequence+'/{}.json'.format(video_name)
with open(path_out_json, 'w') as outfile:
json.dump(video_info, outfile)
if len(video_info['data']) == 0:
print('{} Can not find pose estimation results.'.format(video))
else:
print('{} pose estimation complete.'.format(video))
return
def json_pack(snippets_dir, video_name, frame_width, frame_height, label='unknown', label_index=-1):
'''
将 视频提取的 *.json数据打包成dict
param
snippets_dir:
video_name:
frame_width:
frame_height:
label:
label_index:
return:
video_info:dict{
data:list(dict{
skeleton:list
pose:2d坐标[x,y]
score:坐标的置信分数
})
label:
label_index:
}
'''
sequence_info = []
p = Path(snippets_dir)
for path in p.glob(video_name + '*.json'):
json_path = str(path)
print(path)
frame_id = int(path.stem.split('_')[-2])
frame_data = {'frame_index': frame_id}
data = json.load(open(json_path))
skeletons = []
for person in data['people']:
score, coordinates = [], []
skeleton = {}
keypoints = person['pose_keypoints_2d']
for i in range(0, len(keypoints), 3):
coordinates += [keypoints[i]/frame_width, keypoints[i + 1]/frame_height]
score += [keypoints[i + 2]]
skeleton['pose'] = coordinates
skeleton['score'] = score
skeletons += [skeleton]
frame_data['skeleton'] = skeletons
sequence_info += [frame_data]
video_info = dict()
video_info['data'] = sequence_info
video_info['label'] = label
video_info['label_index'] = label_index
return video_infopreprocess.py
#!/usr/bin/env python
# coding:gbk
import os
from utils.skeleton import extract_skeleon
def PreProcess(path_data, path_resize, path_out):
'''
处理指定路径下的视频数据集，得到调整大小后的视频和提取的骨骼数据
param
path_data: 原视频数据路径
path_resize: 调整大小后的视频路径
path_out: 提取的骨骼数据路径
'''
video_list = os.listdir(path_data)
video_num = len(video_list)
for process_index in range(video_num):
action = video_list[process_index]
label = action
index = process_index
## 1. 将视频调整到指定大小(默认340x256)
path_data_action = '{}/{}'.format(path_data, action)
path_resize_action = '{}/{}'.format(path_resize,action)
# if not os.path.exists(path_resize_action):
# os.makedirs(path_resize_action)
# resize(path_data_action, path_resize_action)
# 2. 利用openpose提取每段视频骨骼点数据
path_out_action = '{}/{}'.format(path_out,action)
if not os.path.exists(path_out_action):
os.makedirs(path_out_action)
extract_skeleon(path_resize_action, path_out_action, label, index)
if __name__ == '__main__':
path_data = 'E:/DataSet/test_data/origin'
path_resize = 'E:/DataSet/test_data/resize'
path_skeleton = 'E:/DataSet/test_data/skeleton'
PreProcess(path_data, path_resize, path_skeleton)
6230
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
