在研究人体模型和人体姿态等问题时，我们往往会使用一些通用的人体姿态表示模型来可视化最终的结果；同时，我们经常在一些高质量的人体运动数据集上进行训练来学习人体运动先验。在这篇文章中，我将会介绍一些我个人研究中会用的一些工具以及使用方法。

🤔 SMPL+H模型

SMPL关节树

用SMPL表示的人体关节坐标探究

📝 AMASS数据集

介绍

使用

import numpy as np
### Read AMASS-ACCAD
npz_data_path = 'B21 s3 - put down box to walk_poses.npz'  #我没打错，它真的叫这个名字
bdata = np.load(npz_data_path)

### bdata中总共包含了6个npy数组
### 分别是：trans.npy gender.npy mocap_framerate.npy betas.npy dmpls.npy poses.npy
print(bdata['mocap_framerate'])
！output: array(120.)
### 说明AMASS数据的帧率是120
num_frames = bdata['trans'].shape[0]
! output: 1289
### 显示该动作序列一共持续了1289帧
trans = bdata['trans'][:]
### trans.npy包含了人体在全局坐标系中的位移
### trans的shape: (1289,3)
root_orient = bdata['poses'][:, :3]   ## shape:(1289,3)
### root_orient表示根关节的旋转弧度
pose_body = bdata['poses'][:, 3:66] 
### pose_body: 身体其他关节的旋转弧度
pose_hand = bdata['poses'][:, 66:]    ## shape:(1289,90)
### 手指关节旋转

trim_data = [trans, root_orient, pose_body, pose_hand]
for i, data_seq in enumerate(trim_data):
		trim_data[i] = data_seq[int(0.1*num_frames):int(0.9*num_frames)]
trans, root_orient, pose_body, pose_hand = trim_data
num_frames = trans.shape[0]
print(num_frames)
! output: 1032

动力学前向计算得到关节位置

### forward pass to get joints using smplh
from body_model.body_model import BodyModel
###  SMPLH需要body_model这个模块
gender = 'female'
bm_path = os.path.join(smplh_path, gender, 'model.npz'). ## smplh_path是下载的SMPLH模型的路径
device = torch.device("cuda" if available else "cpu")
bm = BodyModel(bm_path=bm_path, num_betas=16, batch_size=num_frames).to(device)  ## num_betas代表SMPL使用的体型参数的个数

npz_data = np.load(bm_path)
### model.npz一共包含了12个子数组，它们分别是：
		# J_regressor_prior.npy
		# f.npy
		# J_regressor.npy
		# kintree_table.npy
		# J.npy
		# weights_prior.npy
		# weights.npy
		# posedirs.npy
		# bs_style.npy
		# v_template.npy
		# shapedirs.npy
		# bs_type.npy
ori_kintree_table = npz_data['kintree_table']  # 2*52
kintree_table = ori_kintree_table[0, :22]  # 22
kintree_table[0] = -1
### kintree_table记录的是人体骨骼关节树中，每一个关节的父节点编号;这里根节点的父节点被设置成-1

pose_body = torch.Tensor(pose_body).to(device)
pose_hand = torch.Tensor(pose_hand).to(device)
betas = torch.Tensor(np.repeat(betas[:NUM_BETAS][np.newaxis], num_frames, axis=0)).to(device) 
root_orient = torch.Tensor(root_orient).to(device)
trans = torch.Tensor(trans).to(device)
### 这里的pose_body, betas, trans等变量是上面的代码里从AMASS数据集中读取到的
body = bm(pose_body=pose_body, pose_hand=pose_hand, betas=betas, root_orient=root_orient, trans=trans)

### 得到body之后，事实上我们已经根据body_model和SMPLH得到了关节位置信息；这些信息都在body这个对象中
cur_joint_seq = body.Jtr.data.cpu()  ## shape: (1032,52,3)

from body_model.utils import SMPL_JOINTS, KEYPT_VERTS
cur_body_joint_seq = cur_joint_seq[: , :len(SMPL_JOINTS), :]  ## shape: (1032,52,3)

### 读取三维人体模型的网格信息
cur_vtx_seq = body.v.data.cpu()  ## shape: (1032,6890,3)

进阶

姿态矫正

接触判断

降采样

### fps是实际的帧率；OUTPUT_FPS是降采样之后的输出帧率
fps_ratio = float(OUT_FPS) / fps
print('Downsamp ratio: %f' % (fps_ratio))
new_num_frames = int(fps_ratio*num_frames)
print('Downsamp num frames: %d' % (new_num_frames))
# print(cur_num_frames)
# print(new_num_frames)
downsamp_inds = np.linspace(0, num_frames-1, num=new_num_frames, dtype=int)
# print(downsamp_inds)

# update data to save
fps = OUT_FPS
num_frames = new_num_frames
contacts = contacts[downsamp_inds]
trans = trans[downsamp_inds]
root_orient = root_orient[downsamp_inds]
pose_body = pose_body[downsamp_inds]
pose_hand = pose_hand[downsamp_inds]
joint_seq = joint_seq[downsamp_inds]

计算某个关节的全局位姿

三维空间中平移量的表示很简单，只需要用一个包含三个元素的向量即可；但是对于旋转量却有很多的表示方式，最常见的当然是用一个3*3的矩阵来表示。但是使用矩阵来表示其实包含了很多冗余信息，旋转这件事其实只需要知道一个旋转轴（3元素向量）+一个角度一共4个值就可以表示了，这种用4个数来表示旋转的表示方式被称为四元数。除此之外，还有用6个数来表示旋转的方法，这种表示方式的好处是可以连续地表示旋转过程。将来我可以另外写一篇博客来讨论这两种表示形式。

def local2global_pose(local_pose, kintree):
    # local_pose: T X J X 3 X 3 
    bs = local_pose.shape[0]

    local_pose = local_pose.view(bs, -1, 3, 3)

    global_pose = local_pose.clone()

    for jId in range(len(kintree)):
        parent_id = kintree[jId]
        if parent_id >= 0:
            global_pose[:, jId] = torch.matmul(global_pose[:, parent_id], global_pose[:, jId])

    return global_pose

local_joint_aa = torch.cat((torch.from_numpy(root_orient).float(), torch.from_numpy(pose_body).float()), dim=1) # T X (3+21*3)
local_joint_aa = local_joint_aa.reshape(-1, 22, 3) # T X 22 X 3  
local_joint_rot_mat = transforms.axis_angle_to_matrix(local_joint_aa) # T X 22 X 3 X 3 
global_joint_rot_mat = local2global_pose(local_joint_rot_mat, kintree_table) # T X 22 X 3 X 3

global_head_trans_diff = global_head_trans[1:, :] - global_head_trans[:-1, :] # (T-1) X 3

global_head_rot_mat_diff = torch.matmul(torch.inverse(global_head_rot_mat[:-1]), global_head_rot_mat[1:]) # (T-1) X 3 X 3

👤强化学习环境Mujoco中的人体姿态表示

SMPL位姿转换成qpos位姿

pose_aa

seq_length = amass_pose.shape[0]
pose_aa = torch.cat((torch.tensor(amass_root_pose), torch.tensor(amass_pose), torch.zeros(seq_length, 6)), dim=-1)

trans

curr_trans = amass_trans

qpos

from mujoco_py import load_model_from_path  # 首先导入mujoco的python接口
from scipy.spatial.transform import Rotation as sRot

### 在使用Mujoco之前需要先编写xml格式的配置文件，然后通过该文件创建humanoid
model_file = '***.xml'
humanoid_model = load_model_from_path(model_file)

### 先把SMPL的骨骼名称定义好
SMPL_BONE_NAMES = ['Pelvis', 'L_Hip', 'R_Hip', 'Torso', 'L_Knee', 'R_Knee', 'Spine', 'L_Ankle', 'R_Ankle', 'Chest', 'L_Toe', 'R_Toe', 'Neck', 
                    'L_Thorax', 'R_Thorax', 'Head', 'L_Shoulder', 'R_Shoulder', 'L_Elbow', 'R_Elbow', 'L_Wrist', 'R_Wrist', 'L_Hand', 'R_Hand']

euler_order = 'ZYX'

### 编写一个函数来输出humanoid各个骨关节在数组中的索引
def get_body_qposaddr(model):
    body_qposaddr = dict()
    for i, body_name in enumerate(model.body_names):
        start_joint = model.body_jntadr[i]
        if start_joint < 0:
            continue
        end_joint = start_joint + model.body_jntnum[i]
        start_qposaddr = model.jnt_qposadr[start_joint]
        if end_joint < len(model.jnt_qposadr):
            end_qposaddr = model.jnt_qposadr[end_joint]
        else:
            end_qposaddr = model.nq
        body_qposaddr[body_name] = (start_qposaddr, end_qposaddr)
    return body_qposaddr

smpl_2_mujoco = [ SMPL_BONE_NAMES.index(q) for q in list(get_body_qposaddr(model).keys()) if q in SMPL_BONE_NAMES]

pose = pose_aa.reshape(-1, 24, 3).reshape(-1, 72)

### 计算每一个关节的旋转矩阵: (T*24,3)->(T,24,4,4)
curr_pose_mat = angle_axis_to_rotation_matrix(pose.reshape(-1, 3)).reshape(pose.shape[0], -1, 4, 4)
curr_spose = sRot.from_matrix(curr_pose_mat[:,:,:3,:3].reshape(-1, 3, 3).numpy())
### 将旋转矩阵转换成欧拉角，注意欧拉角的顺序是Z轴->Y轴->X轴
### （T*24,3,3）->(T,72)
curr_spose_euler = curr_spose.as_euler(euler_order, degrees=False).reshape(curr_pose_mat.shape[0], -1)
### 将欧拉角按照Humanoid中骨骼树的先后顺序进行重新排序
curr_spose_euler = curr_spose_euler.reshape(-1, 24, 3)[:, smpl_2_mujoco, :].reshape(-1, 72)
### 将根关节的旋转矩阵转化成四元数
root_quat = rotation_matrix_to_quaternion(curr_pose_mat[:,0,:3,:])
### 将根关节平移向量trans, 根关节的旋转四元数和各个关节的旋转欧拉角进行拼接
### 得到适用于mujoco的qpos格式的人体位姿
curr_qpos = np.concatenate((trans, root_quat, curr_spose_euler[:,3:]), axis = 1)

beta

beta = betas_amass[:10]

expert

将其余参数以及数组打包

smpl_seq_data[dest_seq_name] = {
                        "pose_aa": pose_aa.numpy(),
                        # "pose_6d":pose_seq_6d.numpy(),
                        "qpos": qpos,
                        'trans': curr_trans,
                        'beta': betas[:10],
                        "seq_name": dest_seq_name,
                        "gender": gender,
                        "expert": expert_res
                    }

qpos位姿转换为SMPL位姿

参考文章

1. EgoBody

2. Kinpoly