关键点平滑算法笔记

目录

Peppa_Pig_Face_Landmark

这个也有：

splev平滑

关键点卡尔曼滤波和低通滤波

alpahpose 是跟踪box的

Peppa_Pig_Face_Landmark

这个算法有关键点跟踪，关键点比较稳，模型60M

self.trace = GroupTrack(cfg['Skps']['Trace'])

这个也有：

https://github.com/ostadabbas/3d-facial-landmark-detection-and-tracking/blob/a1f3453d0e454f5f08a6e90ff354cb69b4be0152/landmark_detection_video.py

splev平滑

import numpy as np
import matplotlib.pyplot as plt
from scipy.interpolate import splrep, splev
 
def nihe():
# 创建一组数据点以模拟原始数据
    x = np.linspace(0, 20, 100)
    y = np.sin(x) + np.random.normal(scale=0.1, size=x.shape)
 
    # 使用 splrep 计算这些数据点的 B-样条曲线的系数
    spl = splrep(x, y)
 
    # 在一个更密集的点集上评估 B-样条曲线
    x_dense = np.linspace(0, 20, 200)
    y_spline = splev(x_dense, spl)
 
    # 绘制原始数据点
    plt.scatter(x, y, label='Original Data', color='red')
 
    # 绘制 B-样条曲线
    plt.plot(x_dense, y_spline, label='B-Spline Curve', color='blue')
 
    plt.legend()
    plt.title('B-Spline Interpolation')
    plt.xlabel('X')
    plt.ylabel('Y')
    plt.show()
 
def smooth():
    import numpy as np
    from scipy.interpolate import splrep, splev
    import matplotlib.pyplot as plt
 
    # 创建一组带有噪声的数据点
    x = np.linspace(0, 10, 100)
    y = np.sin(x) + np.random.normal(scale=0.1, size=x.shape)
 
    # 计算 B-样条曲线的系数，调整平滑参数 s 控制平滑程度
    spl = splrep(x, y, s=1)  # 增加 s 值以增加平滑程度
 
    # 在原始数据点上评估 B-样条曲线，实现平滑
    y_smooth = splev(x, spl)
 
    # 绘制原始数据和平滑后的曲线
    plt.scatter(x, y, label='Original Data', color='red', s=10)
    plt.plot(x, y_smooth, label='Smoothed Curve', color='blue')
    plt.legend()
    plt.title('Data Smoothing with B-Spline')
    plt.show()

关键点卡尔曼滤波和低通滤波

import numpy as np
 
import matplotlib.pyplot as plt
 
# rShldrBend 0, rForearmBend 1, rHand 2, rThumb2 3, rMid1 4,
# lShldrBend 5, lForearmBend 6, lHand 7, lThumb2 8, lMid1 9,
# lEar 10, lEye 11, rEar 12, rEye 13, Nose 14,
# rThighBend 15, rShin 16, rFoot 17, rToe 18,
# lThighBend 19,  lShin 20, lFoot 21, lToe 22,
# abdomenUpper 23, hip 24, head 25, neck 26, spine 27
parent =[26,0,1,2,2, 26,5,6,7,7, 11,14,13,14,25, 24,15,16,17, 24,19,20,21, 24,-1,26,27,23]
jointnum = 28
dataOri = np.reshape(np.loadtxt("./unity_data/record.txt"),(-1,jointnum,3))
framenum = dataOri.shape[0]
 
datakalman = np.zeros_like(dataOri)
datafinal = np.zeros_like(dataOri)
 
# kalman filter parameters
KalmanParamQ = 0.001
KalmanParamR = 0.0015
K = np.zeros((jointnum,3),dtype=np.float32)
P = np.zeros((jointnum,3),dtype=np.float32)
X = np.zeros((jointnum,3),dtype=np.float32)
# low pass filter parameters
PrevPose3D = np.zeros((6,jointnum,3),dtype=np.float32)
for idx in range(framenum):
    currdata = np.squeeze(dataOri[idx])
    smooth_kps = np.zeros((jointnum,3),dtype=np.float32)
    '''
    kalman filter
    '''
    for i in range(jointnum):
        K[i] = (P[i] + KalmanParamQ) / (P[i] + KalmanParamQ + KalmanParamR)
        P[i] = KalmanParamR * (P[i] + KalmanParamQ) / (P[i] + KalmanParamQ + KalmanParamR)
    for i in range(jointnum):
        smooth_kps[i] = X[i] + (currdata[i] - X[i])*K[i]
        X[i] = smooth_kps[i]
 
    datakalman[idx] = smooth_kps # record kalman result
 
    '''
    low pass filter
    '''    
    LowPassParam = 0.1
    PrevPose3D[0] = smooth_kps
    for j in range(1,6):
        PrevPose3D[j] = PrevPose3D[j] * LowPassParam + PrevPose3D[j - 1] * (1.0 - LowPassParam)
    datafinal[idx] = PrevPose3D[5] # record kalman+low pass result.
 
'''
visualization
'''
# visualize original/pythonsmooth/unitysmooth z coordinates for the joint 17
dataSmooth = np.reshape(np.loadtxt("./unity_data/recordSmooth.txt"),(-1,28,3))
x1 = np.arange(0,framenum,1)
y1 = np.squeeze(dataOri[:,17,1])
x2 = np.arange(0,dataSmooth.shape[0],1)
y2 = np.squeeze(dataSmooth[:,17,1])
plt.plot(x1,y1)
plt.plot(x2,y2,linewidth = 1)
plt.plot(x1,datakalman[:,17,1],linewidth=2,linestyle=":")
plt.plot(x1,datafinal[:,17,1],linewidth=1,linestyle="-.")
plt.legend(["original","unitySmooth","kalman","kalman+lowpass"])
plt.show()
 
# visualize the skeleton animation
fig = plt.figure()
plt.ion()
data = datafinal#dataOri
for i in range(framenum):   
    fig.clf() 
    kps = np.squeeze(data[i,...])    
    ax = fig.gca(projection='3d')
    ax.xaxis.set_label_text(label="x")
    ax.yaxis.set_label_text(label="y")
    ax.zaxis.set_label_text(label="z")
    ax.set_xlim(np.min(data[...,0]),np.max(data[...,0]))
    ax.set_ylim(np.min(data[...,2]),np.max(data[...,2]))
    ax.set_zlim(np.min(data[...,1]),np.max(data[...,1]))
    ax.view_init(elev=26., azim=70)    
    ax.scatter3D(kps[:,0],-kps[:,2],kps[:,1],'red')
    for i in range(28):
        if(parent[i]!=-1):
            ax.plot3D(kps[[i,parent[i]],0], -kps[[i,parent[i]],2], kps[[i,parent[i]],1], 'gray')
    plt.pause(0.02)    
plt.ioff()
plt.show()

alpahpose 是跟踪box的

https://github.com/MVIG-SJTU/AlphaPose/blob/c60106d19afb443e964df6f06ed1842962f5f1f7/trackers/tracker_api.py