运动识别之HOJ3D和HMM

http://cvrc.ece.utexas.edu/Publications/Xia_HAU3D12.pdf

 

View Invariant Human Action Recognition Using Histograms of 3D Joints

---

The  HOJ3D computed from the action depth sequences are  reprojected using LDA and then clustered into k posture  visual words, which  represent  the  prototypical  poses  of  actions.  The temporal evolutions of  those  visual  words  are modeled by discrete hidden Markov models (HMMs).

 

 

---

特征定义

 

In this representation, the 3D space is partitioned into n bins using a modified spherical coordinate system. We manually select 12 informative joints to build a compact representation of human posture. To make our representation robust against minor posture variation, votes of 3D skeletal joints are cast into neighboring bins using a Gaussian weight function.

 

we acquire  the  3D locations of 20 skeletal joints  which  comprise  hip  center,  spine,  shoulder  center, head, L/ R shoulder, L/ R elbow, L/ R wrist, L/ R hand, L/ R hip, L/  R knee, L/  R angle and L/ R foot. 

 

we  compute  our  histogram based  representation of postures  from 12 of the 20 joints, including  head, L/  R elbow, L/  R hands, L/  R knee, L/ R feet, hip center  and  L/  R hip. We take the hip center  as the center of  the  reference  coordinate  system,  and  define  the x-direction according to L/ R hip. The rest 9 joints are used to compute the 3D spatial histogram.

 

要达到视不变（不同视角下相同姿态正确归类）：We achieve this by  aligning  our  spherical  coordinates  with  the  person’s specific  direction。We  define  the center of the spherical coordinates as the  hip center  joint.Define  the  horizontal  reference  vector  α  to  be  the  vector from the left hip center to  the  right hip center  projected on the horizontal plane (parallel to the ground), and the zenith reference vector θ as the vector that is perpendicular to the ground plane and passes through the coordinate center.

 

partition the 3D space into n  bins

The  inclination angle is divided into 7 bins from  the zenith vector θ: [0, 15], [15,  45], [45, 75], [105, 135], [165, 180]

 

Our HOJ3D descriptor is computed by casting the rest 9 joints  into  the  corresponding  spatial  histogram  bins.

To  make  the representation robust against minor errors of joint locations, we vote the 3D bins using a Gaussian weight function:

 

For each  joint,  we  only  vote  over  the  bin  it  is  in  and  the  8 neighboring bins. We calculate the probabilistic voting on θ and  α  separately since they are independent (see Fig. 4). The probabilistic voting for each of the 9 bins is the product of the probability on α  direction and θ direction. Let the joint 

location  be The  vote  of  a  joint  location  to  bin  is 

 

输入为20*3（20个关节点，xyz3维空间坐标），输出为84位HOJ3D特征

 

特征为84维向量，水平方向12，垂直方向7

1，12个关节点局部坐标的计算：1，根据L_HIP和R_HIP的连线方向计算转换后的坐标 ； 2，计算相对于HIP_CENTER的坐标

2，之后计算两个偏转角 vector  α 和 vector θ

3，在每个关节所属的bin中的8个邻域内，按双方向的单高斯分布乘积投票

 

---

特征降维

 

Linear  discriminant  analysis  (LDA)  is  performed  to extract the  dominant  features. 

降维的目的是得到区分度更大的9个维度信息

 

输入为84维HOJ3D特征，输出为9维降维特征

---

特征聚类

 

We cluster the vectors into K clusters (a K-word vocabulary) using K-means. Then each posture is represented as a  single number of a  visual word.

聚类是为了减少观察特征表示，训练阶段需要把所有观测数据（所有动作，每一个动作包含若干帧，每帧的20个骨骼节点经过LDA降维成9）在9维空间中聚类，可以得到25个聚类中心的坐标（9维），依次标号

在识别阶段，将LDA之后的特征，分配到最近邻的聚类中心，记录其标号，作为HMM的输入参数

 

训练阶段，输入为所有动作的9维特征，输出为25个聚类中心

识别阶段，输入为每一帧的动作特征（9维），输出为距其最近的聚类中心的标号

 

---

动作识别

 

the HMM gives a state based representation for each action.  After forming the models for each activity, we take an action sequence and  calculate  its  probability  of  a  modelfor  the observation  sequence, for every model, which  can be solved using the forward algorithm. Then we classify the 

action as the one which has the largest posterior probability.

 

训练阶段，输入为每一类动作的时序标号，输出为HMM模型参数

识别阶段，输入为某一动作的时序标号，输出为前向概率即模型匹配度最大的动作模型 —— 识别结束

 

---

---

 

来自为知笔记(Wiz)