文章目录
- 前言
- 一、所需准备
- 1、fork原文
- 2、全部运行
- 3、更新库文件
- 二、正式使用
- 1、骨骼关键点检测
- 2、将皮影素材映射到骨骼关键点上
- 3、让皮影动起来
- 4、视频效果
- 三、总结
前言
PaddleHub是飞桨生态下的预训练模型的管理工具,旨在让飞桨生态下的开发者更便捷地享受到大规模预训练模型的价值。用户可以通过PaddleHub便捷地获取飞桨生态下的预训练模型,结合Fine-tune API快速完成迁移学习到应用部署的全流程工作,让预训练模型能更好服务于用户特定场景的应用。
当前PaddleHub已经可以支持文本、图像和视频三大类主流方向,为用户准备了大量高质量的预训练模型,可以满足用户各种应用场景的任务需求,包括但不限于词法分析、情感分析、图像分类、目标检测、视频分类等经典任务。
在PaddleHub获取到人体骨骼关键点模型之后,就可以对这些关键点进行连接,从而形成了人体姿态。然后我们将皮影素材映射到人体姿态身上,让皮影跟随人体姿态进行运动,就达到“皮影戏”的效果。
一、所需准备
本项目可以通过两种途径实现,一种是线下,一种是线上。线下需要自己装在环境到电脑中,考虑到python版本之间的不兼容,而且自己电脑配置较低,因而为了防止出现错误,本文所使用的是通过线上运行的方式。
1、fork原文
通过fork原文,我们在network会备份出和作者一样的环境
如果还是想要自己在本地运行,建议打开终端查看python库文件版本,然后下载和这个示例里一样的版本
pip list
通过pip list可以看到paddlepaddle和paddlehub等库文件的版本
2、全部运行
看着原文的示例,应该是成功过一次的,我们来全部运行一下
可以看到,这里提示报错了,根据错误提示,需要更新paddlepaddle库
3、更新库文件
paddlepaddle官方文档 根据官方文档提示,我们知道paddlepaddle对环境是有要求的
python版本 | 3.5.1+/3.6/3.7/3.8 |
pip版本 | 20.2.2+ |
处理器架构 | 第一行输出的是”64bit”,第二行输出的是”x86_64”、”x64”或”AMD64”即可 |
通过python -V我们可以查看当前python版本
python -V
通过pip show pip,我们可以查看到当前的pip版本
pip show pip
通过python -m pip install -U pip,更新pip版本
python -m pip install -U pip
通过pip install -U your module更新库,这里的库就是paddlepaddle
pip install -U paddlepaddle
更新完毕后,再次使用pip list指令查看,发现更新成功
二、正式使用
1、骨骼关键点检测
上传自己想要检测的图片,运行代码
import os
import cv2
import paddlehub as hub
import matplotlib.pyplot as plt
from matplotlib.image import imread
import numpy as np
%matplotlib inline
def show_img(img_path, size=8):
'''
文件读取图片显示
'''
im = imread(img_path)
plt.figure(figsize=(size,size))
plt.axis("off")
plt.imshow(im)
def img_show_bgr(image,size=8):
'''
cv读取的图片显示
'''
image=cv2.cvtColor(image,cv2.COLOR_BGR2RGB)
plt.figure(figsize=(size,size))
plt.imshow(image)
plt.axis("off")
plt.show()
#通过代码获取图片中的结果
pose_estimation = hub.Module(name="human_pose_estimation_resnet50_mpii")
result = pose_estimation.keypoint_detection(paths=['work/imgs/body01.jpg'], visualization=True, output_dir="work/output_pose/")
result
由代码可以知道,检测结果会输出到output_pose这个文件夹中,让我们来看一下检测结果如何,我们可以看到元原来的图片中多出了一些身体上的检测点,可以知道这是成功检测到了骨骼关键点了
2、将皮影素材映射到骨骼关键点上
皮影素材
运行映射python程序
def get_true_angel(value):
'''
转转得到角度值
'''
return value/np.pi*180
def get_angle(x1, y1, x2, y2):
'''
计算旋转角度
'''
dx = abs(x1- x2)
dy = abs(y1- y2)
result_angele = 0
if x1 == x2:
if y1 > y2:
result_angele = 180
else:
if y1!=y2:
the_angle = int(get_true_angel(np.arctan(dx/dy)))
if x1 < x2:
if y1>y2:
result_angele = -(180 - the_angle)
elif y1<y2:
result_angele = -the_angle
elif y1==y2:
result_angele = -90
elif x1 > x2:
if y1>y2:
result_angele = 180 - the_angle
elif y1<y2:
result_angele = the_angle
elif y1==y2:
result_angele = 90
if result_angele<0:
result_angele = 360 + result_angele
return result_angele
def rotate_bound(image, angle, key_point_y):
'''
旋转图像,并取得关节点偏移量
'''
#获取图像的尺寸
(h,w) = image.shape[:2]
#旋转中心
(cx,cy) = (w/2,h/2)
# 关键点必须在中心的y轴上
(kx,ky) = cx, key_point_y
d = abs(ky - cy)
#设置旋转矩阵
M = cv2.getRotationMatrix2D((cx,cy), -angle, 1.0)
cos = np.abs(M[0,0])
sin = np.abs(M[0,1])
# 计算图像旋转后的新边界
nW = int((h*sin)+(w*cos))
nH = int((h*cos)+(w*sin))
# 计算旋转后的相对位移
move_x = nW/2 + np.sin(angle/180*np.pi)*d
move_y = nH/2 - np.cos(angle/180*np.pi)*d
# 调整旋转矩阵的移动距离(t_{x}, t_{y})
M[0,2] += (nW/2) - cx
M[1,2] += (nH/2) - cy
return cv2.warpAffine(image,M,(nW,nH)), int(move_x), int(move_y)
def get_distences(x1, y1, x2, y2):
return ((x1-x2)**2 + (y1-y2)**2)**0.5
def append_img_by_sk_points(img, append_img_path, key_point_y, first_point, second_point, append_img_reset_width=None,
append_img_max_height_rate=1, middle_flip=False, append_img_max_height=None):
'''
将需要添加的肢体图片进行缩放
'''
append_image = cv2.imdecode(np.fromfile(append_img_path, dtype=np.uint8), cv2.IMREAD_UNCHANGED)
# 根据长度进行缩放
sk_height = int(get_distences(first_point[0], first_point[1], second_point[0], second_point[1])*append_img_max_height_rate)
# 缩放制约
if append_img_max_height:
sk_height = min(sk_height, append_img_max_height)
sk_width = int(sk_height/append_image.shape[0]*append_image.shape[1]) if append_img_reset_width is None else int(append_img_reset_width)
if sk_width <= 0:
sk_width = 1
if sk_height <= 0:
sk_height = 1
# 关键点映射
key_point_y_new = int(key_point_y/append_image.shape[0]*append_image.shape[1])
# 缩放图片
append_image = cv2.resize(append_image, (sk_width, sk_height))
img_height, img_width, _ = img.shape
# 是否根据骨骼节点位置在 图像中间的左右来控制是否进行 左右翻转图片
# 主要处理头部的翻转, 默认头部是朝左
if middle_flip:
middle_x = int(img_width/2)
if first_point[0] < middle_x and second_point[0] < middle_x:
append_image = cv2.flip(append_image, 1)
# 旋转角度
angle = get_angle(first_point[0], first_point[1], second_point[0], second_point[1])
append_image, move_x, move_y = rotate_bound(append_image, angle=angle, key_point_y=key_point_y_new)
app_img_height, app_img_width, _ = append_image.shape
zero_x = first_point[0] - move_x
zero_y = first_point[1] - move_y
(b, g, r) = cv2.split(append_image)
for i in range(0, r.shape[0]):
for j in range(0, r.shape[1]):
if 230>r[i][j]>200 and 0<=zero_y+i<img_height and 0<=zero_x+j<img_width:
img[zero_y+i][zero_x+j] = append_image[i][j]
return img
body_img_path_map = {
"right_hip" : "./work/shadow_play_material/right_hip.jpg",
"right_knee" : "./work/shadow_play_material/right_knee.jpg",
"left_hip" : "./work/shadow_play_material/left_hip.jpg",
"left_knee" : "./work/shadow_play_material/left_knee.jpg",
"left_elbow" : "./work/shadow_play_material/left_elbow.jpg",
"left_wrist" : "./work/shadow_play_material/left_wrist.jpg",
"right_elbow" : "./work/shadow_play_material/right_elbow.jpg",
"right_wrist" : "./work/shadow_play_material/right_wrist.jpg",
"head" : "./work/shadow_play_material/head.jpg",
"body" : "./work/shadow_play_material/body.jpg"
}
def get_combine_img(img_path, pose_estimation=pose_estimation, body_img_path_map=body_img_path_map, backgroup_img_path= 'work/background.jpg'):
'''
识别图片中的关节点,并将皮影的肢体进行对应,最后与原图像拼接后输出
'''
result = pose_estimation.keypoint_detection(paths=[img_path])
image=cv2.imread(img_path)
# 背景图片
backgroup_image = cv2.imread(backgroup_img_path)
image_flag = cv2.resize(backgroup_image, (image.shape[1], image.shape[0]))
# 最小宽度
min_width = int(get_distences(result[0]['data']['head_top'][0], result[0]['data']['head_top'][1],
result[0]['data']['upper_neck'][0], result[0]['data']['upper_neck'][1])/3)
#右大腿
append_img_reset_width = max(int(get_distences(result[0]['data']['pelvis'][0], result[0]['data']['pelvis'][1],
result[0]['data']['left_hip'][0], result[0]['data']['right_hip'][1])*1.6), min_width)
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['right_hip'], key_point_y=10, first_point=result[0]['data']['right_hip'],
second_point=result[0]['data']['right_knee'], append_img_reset_width=append_img_reset_width)
# 右小腿
append_img_reset_width = max(int(get_distences(result[0]['data']['pelvis'][0], result[0]['data']['pelvis'][1],
result[0]['data']['left_hip'][0], result[0]['data']['right_hip'][1])*1.5), min_width)
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['right_knee'], key_point_y=10, first_point=result[0]['data']['right_knee'],
second_point=result[0]['data']['right_ankle'], append_img_reset_width=append_img_reset_width)
# 左大腿
append_img_reset_width = max(int(get_distences(result[0]['data']['pelvis'][0], result[0]['data']['pelvis'][1],
result[0]['data']['left_hip'][0], result[0]['data']['left_hip'][1])*1.6), min_width)
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['left_hip'], key_point_y=0, first_point=result[0]['data']['left_hip'],
second_point=result[0]['data']['left_knee'], append_img_reset_width=append_img_reset_width)
# 左小腿
append_img_reset_width = max(int(get_distences(result[0]['data']['pelvis'][0], result[0]['data']['pelvis'][1],
result[0]['data']['left_hip'][0], result[0]['data']['left_hip'][1])*1.5), min_width)
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['left_knee'], key_point_y=10, first_point=result[0]['data']['left_knee'],
second_point=result[0]['data']['left_ankle'], append_img_reset_width=append_img_reset_width)
# 右手臂
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['left_elbow'], key_point_y=25, first_point=result[0]['data']['right_shoulder'],
second_point=result[0]['data']['right_elbow'], append_img_max_height_rate=1.2)
# 右手肘
append_img_max_height = int(get_distences(result[0]['data']['right_shoulder'][0], result[0]['data']['right_shoulder'][1],
result[0]['data']['right_elbow'][0], result[0]['data']['right_elbow'][1])*1.6)
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['left_wrist'], key_point_y=10, first_point=result[0]['data']['right_elbow'],
second_point=result[0]['data']['right_wrist'], append_img_max_height_rate=1.5,
append_img_max_height=append_img_max_height)
# 左手臂
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['right_elbow'], key_point_y=25, first_point=result[0]['data']['left_shoulder'],
second_point=result[0]['data']['left_elbow'], append_img_max_height_rate=1.2)
# 左手肘
append_img_max_height = int(get_distences(result[0]['data']['left_shoulder'][0], result[0]['data']['left_shoulder'][1],
result[0]['data']['left_elbow'][0], result[0]['data']['left_elbow'][1])*1.6)
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['right_wrist'], key_point_y=10, first_point=result[0]['data']['left_elbow'],
second_point=result[0]['data']['left_wrist'], append_img_max_height_rate=1.5,
append_img_max_height=append_img_max_height)
# 头
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['head'], key_point_y=10, first_point=result[0]['data']['head_top'],
second_point=result[0]['data']['upper_neck'], append_img_max_height_rate=1.2, middle_flip=True)
# 身体
append_img_reset_width = max(int(get_distences(result[0]['data']['left_shoulder'][0], result[0]['data']['left_shoulder'][1],
result[0]['data']['right_shoulder'][0], result[0]['data']['right_shoulder'][1])*1.2), min_width*3)
image_flag = append_img_by_sk_points(image_flag, body_img_path_map['body'], key_point_y=20, first_point=result[0]['data']['upper_neck'],
second_point=result[0]['data']['pelvis'], append_img_reset_width=append_img_reset_width, append_img_max_height_rate=1.2)
result_img = np.concatenate((image, image_flag), axis=1)
return result_img
pos_img_path = 'work/output_pose/body01.jpg'
result_img = get_combine_img(pos_img_path, pose_estimation, body_img_path_map)
img_show_bgr(result_img, size=10)
可以看到映射结果如下,还是挺贴合的
3、让皮影动起来
在这一部分中,我们需要用到OpenCV对视频进行裁切,将视频逐帧分切成一张张的图片,然后再对每一张图片进行骨骼关键点检测并将皮影素材映射上去,最后还原成为视频。
如下是OpenCV拆分的结果,共3036张图片
然后这个是最耗时间的,将一张张图片映射皮影素材上去,花了一个小时
我们需要将自己想要的视频上传上去,视频越短越好,帧率越低越好,因为程序运行时间与帧率和视频长度成正比
运行以下python程序
# 素材图片位置
input_video = 'work/001.mp4'
def transform_video_to_image(video_file_path, img_path):
'''
将视频中每一帧保存成图片
'''
video_capture = cv2.VideoCapture(video_file_path)
fps = video_capture.get(cv2.CAP_PROP_FPS)
count = 0
while(True):
ret, frame = video_capture.read()
if ret:
cv2.imwrite(img_path + '%d.jpg' % count, frame)
count += 1
else:
break
video_capture.release()
print('视频图片保存成功, 共有 %d 张' % count)
return fps
# 将视频中每一帧保存成图片
fps = transform_video_to_image(input_video, 'work/mp4_img/')
def analysis_pose(input_frame_path, output_frame_path, is_print=True):
'''
分析图片中的人体姿势, 并转换为皮影姿势,输出结果
'''
file_items = os.listdir(input_frame_path)
file_len = len(file_items)
for i, file_item in enumerate(file_items):
if is_print:
print(i+1,'/', file_len, ' ', os.path.join(output_frame_path, file_item))
combine_img = get_combine_img(os.path.join(input_frame_path, file_item))
cv2.imwrite(os.path.join(output_frame_path, file_item), combine_img)
# 分析图片中的人体姿势, 并转换为皮影姿势,输出结果
analysis_pose('work/mp4_img/', 'work/mp4_img_analysis/', is_print=False)
def combine_image_to_video(comb_path, output_file_path, fps=30, is_print=False):
'''
合并图像到视频
'''
fourcc = cv2.VideoWriter_fourcc(*'MP4V')
file_items = os.listdir(comb_path)
file_len = len(file_items)
# print(comb_path, file_items)
if file_len > 0 :
temp_img = cv2.imread(os.path.join(comb_path, file_items[0]))
img_height, img_width = temp_img.shape[0], temp_img.shape[1]
out = cv2.VideoWriter(output_file_path, fourcc, fps, (img_width, img_height))
for i in range(file_len):
pic_name = os.path.join(comb_path, str(i)+".jpg")
if is_print:
print(i+1,'/', file_len, ' ', pic_name)
img = cv2.imread(pic_name)
out.write(img)
out.release()
# 合并图像到视频
combine_image_to_video('work/mp4_img_analysis/', 'work/mp4_analysis.mp4', fps)
# 添加音频 mp4_analysis_result.mp4为最终输出文件
! ffmpeg -i work/mp4_analysis.mp4 -i work/001.mp4 -c:v copy -c:a copy work/mp4_analysis_result.mp4 -y
4、视频效果
三、总结
通过本次对paddlepaddle和paddlehub的使用,认识到了百度飞桨平台,学习了python的深度学习,在此再次感谢原作者的开源项目。