2 扩展卡尔曼滤波EKF
在上一节中,我们了解到了卡尔曼滤波的计算公式。卡尔曼滤波基于线性系统的假设,如果运动模型或者观测模型不能用线性系统来表示(大部分现实问题都无法遵从线性系统的假设),那么我们仍然可以使用卡尔曼滤波的思想,只不过我们使用一阶雅克比矩阵来代替状态转移矩阵来进行计算(证明略),这就是扩展卡尔曼滤波EKF:
Localization process using Extendted Kalman Filter:EKF is
=== Predict ===
x P r e d = F x t + B u t x_{Pred} = Fx_t+Bu_t xPred=Fxt+But
P P r e d = J F P t J F T + Q P_{Pred} = J_FP_t J_F^T + Q PPred=JFPtJFT+Q
=== Update ===
z P r e d = H x P r e d z_{Pred} = Hx_{Pred} zPred=HxPred
y = z − z P r e d y = z - z_{Pred} y=z−zPred
S = J H P P r e d . J H T + R S = J_H P_{Pred}.J_H^T + R S=JHPPred.JHT+R
K = P P r e d . J H T S − 1 K = P_{Pred}.J_H^T S^{-1} K=PPred.JHTS−1
x t + 1 = x P r e d + K y x_{t+1} = x_{Pred} + Ky xt+1=xPred+Ky
P t + 1 = ( I − K J H ) P P r e d P_{t+1} = ( I - K J_H) P_{Pred} Pt+1=(I−KJH)PPred
注:首先,在以上公式中, x P r e d = F x t + B u t x_{Pred} = Fx_t+Bu_t xPred=Fxt+But的一种更普遍的表达方式是: x P r e d = g ( x t , u t ) x_{Pred} = g(x_t,u_t) xPred=g(xt,ut),以表现运动模型的非线性。在此继续采用 F x t + B u t Fx_t+Bu_t Fxt+But矩阵相乘的形式是为了对应下面的python代码,实际上我们可以看到在代码中运动模型还是非线性的(原因是 B B B矩阵包含了状态量)。同样地, z P r e d = H x P r e d z_{Pred} = Hx_{Pred} zPred=HxPred 的一种更普遍的表达方式是 z P r e d = h ( x P r e d ) z_{Pred} = h(x_{Pred}) zPred=h(xPred),在此采用矩阵相乘形式是为了对应下面的python代码。
上面公式中, J F J_F JF代表了运动模型的雅克比矩阵, J H J_H JH代表了观测模型的雅克比矩阵。扩展卡尔曼滤波
下面通过一个python实例来展示EKF的核心概念。(完整代码见原链接,有中文注释的附在了最后)
图中是一个应用Extended Kalman Filter(EKF)做传感器融合定位的实例。
蓝色线是轨迹真值,黑色线是“航迹推测”得出的轨迹(航迹推测指的是单纯凭借速度信息推测位置的方法,会将速度信息的误差包含在位置中)
绿色点是观测值(ex. GPS), 红色线是经过EKF滤波后的轨迹。
红色椭圆代表EKF输出的机器人状态的实时协方差。
源代码链接: PythonRobotics/extended_kalman_filter.py at master · AtsushiSakai/PythonRobotics
滤波器设计
在这个实例中,二维机器人具有一个四个元素的状态向量:
x t = [ x t , y t , ϕ t , v t ] \textbf{x}_t=[x_t, y_t, \phi_t, v_t] xt=[xt,yt,ϕt,vt]
x, y 是 2D x-y 位置, ϕ \phi ϕ 是角度, v 是线速度.
在代码中, 用"xEst" 代表状态向量. code
P t P_t Pt 代表状态的协方差矩阵,
Q Q Q 运动模型噪声的协方差矩阵,
R R R 观测模型噪声的协方差矩阵,具体见代码注释(注意,在这个实例中, Q Q Q和 R R R都是固定的)
机器人装备有线速度传感器和角速度传感器,因此运动模型的控制输入向量可以用线速度和角速度表示:
u t = [ v t , ω t ] \textbf{u}_t=[v_t, \omega_t] ut=[vt,ωt]
机器人还有 GNSS 传感器(相当于GPS),意味着机器人可以获得x-y位置定位的观测值:
z t = [ x t , y t ] \textbf{z}_t=[x_t,y_t] zt=[xt,yt]
别忘了,为了模拟实际情况,控制输入向量和观测值向量都应该带有噪声。
在源代码中, “observation” 函数通过在理论值上人为加随机噪声来模拟实际信号code
def observation(xTrue, xd, u):
"""
执行仿真过程,不是EKF的一部分
"""
# 轨迹真值
xTrue = motion_model(xTrue, u)
# add noise to gps x-y
z = observation_model(xTrue) + GPS_NOISE @ np.random.randn(2, 1)
# add noise to input
ud = u + INPUT_NOISE @ np.random.randn(2, 1)
# 航迹推测得出的轨迹:
xd = motion_model(xd, ud)
return xTrue, z, xd, ud
运动模型
机器人的运动模型为:
x ˙ = v c o s ( ϕ ) \dot{x} = vcos(\phi) x˙=vcos(ϕ)
y ˙ = v s i n ( ( ϕ ) \dot{y} = vsin((\phi) y˙=vsin((ϕ)
ϕ ˙ = ω \dot{\phi} = \omega ϕ˙=ω
运动模型方程:
x t + 1 = F x t + B u t \textbf{x}_{t+1} = F\textbf{x}_t+B\textbf{u}_t xt+1=Fxt+But
其中,
F = [ 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 ] \begin{aligned} F= \begin{bmatrix} 1 & 0 & 0 & 0\\ 0 & 1 & 0 & 0\\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 0 \\ \end{bmatrix} \end{aligned} F=⎣⎢⎢⎡1000010000100000⎦⎥⎥⎤
B = [ c o s ( ϕ ) d t 0 s i n ( ϕ ) d t 0 0 d t 1 0 ] \begin{aligned} B= \begin{bmatrix} cos(\phi)dt & 0\\ sin(\phi)dt & 0\\ 0 & dt\\ 1 & 0\\ \end{bmatrix} \end{aligned} B=⎣⎢⎢⎡cos(ϕ)dtsin(ϕ)dt0100dt0⎦⎥⎥⎤
d t dt dt 是时间步长。
注意实际上矩阵 B B B包含了状态向量(机器人角度),这不再是一个单纯的线性模型。
这部分的源代码: code
运动模型的雅克比矩阵(也就是按照状态向量中的各个元素依次彼此求一阶偏导)为:
J F = [ d x d x d x d y d x d ϕ d x d v d y d x d y d y d y d ϕ d y d v d ϕ d x d ϕ d y d ϕ d ϕ d ϕ d v d v d x d v d y d v d ϕ d v d v ] \begin{aligned} J_F= \begin{bmatrix} \frac{dx}{dx}& \frac{dx}{dy} & \frac{dx}{d\phi} & \frac{dx}{dv}\\ \frac{dy}{dx}& \frac{dy}{dy} & \frac{dy}{d\phi} & \frac{dy}{dv}\\ \frac{d\phi}{dx}& \frac{d\phi}{dy} & \frac{d\phi}{d\phi} & \frac{d\phi}{dv}\\ \frac{dv}{dx}& \frac{dv}{dy} & \frac{dv}{d\phi} & \frac{dv}{dv}\\ \end{bmatrix} \end{aligned} JF=⎣⎢⎢⎢⎡dxdxdxdydxdϕdxdvdydxdydydydϕdydvdϕdxdϕdydϕdϕdϕdvdvdxdvdydvdϕdvdv⎦⎥⎥⎥⎤
= [ 1 0 − v s i n ( ϕ ) d t c o s ( ϕ ) d t 0 1 v c o s ( ϕ ) d t s i n ( ϕ ) d t 0 0 1 0 0 0 0 1 ] \begin{aligned} = \begin{bmatrix} 1& 0 & -v sin(\phi)dt & cos(\phi)dt\\ 0 & 1 & v cos(\phi)dt & sin(\phi) dt\\ 0 & 0 & 1 & 0\\ 0 & 0 & 0 & 1\\ \end{bmatrix} \end{aligned} =⎣⎢⎢⎡10000100−vsin(ϕ)dtvcos(ϕ)dt10cos(ϕ)dtsin(ϕ)dt01⎦⎥⎥⎤
def motion_model(x, u):
"""
运动模型
"""
F = np.array([[1.0, 0, 0, 0],
[0, 1.0, 0, 0],
[0, 0, 1.0, 0],
[0, 0, 0, 0]])
# 注意:在这里B矩阵中耦合了状态向量x,因此并不是简单的线性模型:
B = np.array([[DT * math.cos(x[2, 0]), 0],
[DT * math.sin(x[2, 0]), 0],
[0.0, DT],
[1.0, 0.0]])
x = F @ x + B @ u
return x
def jacob_f(x, u):
"""
Jacobian of Motion Model
motion model
x_{t+1} = x_t+v*dt*cos(yaw)
y_{t+1} = y_t+v*dt*sin(yaw)
yaw_{t+1} = yaw_t+omega*dt
v_{t+1} = v{t}
so
dx/dyaw = -v*dt*sin(yaw)
dx/dv = dt*cos(yaw)
dy/dyaw = v*dt*cos(yaw)
dy/dv = dt*sin(yaw)
"""
yaw = x[2, 0]
v = u[0, 0]
jF = np.array([
[1.0, 0.0, -DT * v * math.sin(yaw), DT * math.cos(yaw)],
[0.0, 1.0, DT * v * math.cos(yaw), DT * math.sin(yaw)],
[0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 0.0, 1.0]])
return jF
观测模型
在这个实例中,机器人可以通过GPS直接获取二维 x-y 坐标。
所以GPS的观测模型是:
z t = H x t \textbf{z}_{t} = H\textbf{x}_t zt=Hxt
其中
H = [ 1 0 0 0 0 1 0 0 ] \begin{aligned} H= \begin{bmatrix} 1 & 0 & 0& 0\\ 0 & 1 & 0& 0\\ \end{bmatrix} \end{aligned} H=[10010000]
(即直接获取某时刻x、y的值,与角度速度无关)
对应的雅克比矩阵:
J H = [ d x d x d x d y d x d ϕ d x d v d y d x d y d y d y d ϕ d y d v ] \begin{aligned} J_H= \begin{bmatrix} \frac{dx}{dx}& \frac{dx}{dy} & \frac{dx}{d\phi} & \frac{dx}{dv}\\ \frac{dy}{dx}& \frac{dy}{dy} & \frac{dy}{d\phi} & \frac{dy}{dv}\\ \end{bmatrix} \end{aligned} JH=[dxdxdxdydydxdydydϕdxdϕdydvdxdvdy]
= [ 1 0 0 0 0 1 0 0 ] \begin{aligned} = \begin{bmatrix} 1& 0 & 0 & 0\\ 0 & 1 & 0 & 0\\ \end{bmatrix} \end{aligned} =[10010000]
def observation_model(x):
H = np.array([
[1, 0, 0, 0],
[0, 1, 0, 0]
])
z = H @ x
return z
def jacob_h():
# Jacobian of Observation Model
jH = np.array([
[1, 0, 0, 0],
[0, 1, 0, 0]
])
return jH
以下是完整程序中文注释版:
"""
Extended kalman filter (EKF) localization sample
author: Atsushi Sakai (@Atsushi_twi)
https://github.com/AtsushiSakai/PythonRobotics/blob/master/Localization/extended_kalman_filter/
"""
import math
import matplotlib.pyplot as plt
import numpy as np
# Covariance for EKF:
# 运动模型协方差:
Q = np.diag([
0.1, # variance of location on x-axis
0.1, # variance of location on y-axis
np.deg2rad(10), # variance of yaw angle
10.0 # variance of velocity
]) ** 2 # predict state covariance
# 观测模型协方差:
R = np.diag([1.0, 1.0]) ** 2 # Observation x,y position covariance
# Simulation parameter
INPUT_NOISE = np.diag([1.0, np.deg2rad(30.0)]) ** 2
GPS_NOISE = np.diag([0.5, 0.5]) ** 2
DT = 0.1 # time tick [s]
SIM_TIME = 50.0 # simulation time [s]
show_animation = True
def calc_input():
v = 1.0 # [m/s]
yawrate = 0.1 # [rad/s]
u = np.array([[v], [yawrate]])
return u
def observation(xTrue, xd, u):
"""
执行仿真过程,不是EKF的一部分
"""
# 轨迹真值
xTrue = motion_model(xTrue, u)
# add noise to gps x-y
z = observation_model(xTrue) + GPS_NOISE @ np.random.randn(2, 1)
# add noise to input
ud = u + INPUT_NOISE @ np.random.randn(2, 1)
# 航迹推测得出的轨迹:
xd = motion_model(xd, ud)
return xTrue, z, xd, ud
def motion_model(x, u):
"""
运动模型
"""
F = np.array([[1.0, 0, 0, 0],
[0, 1.0, 0, 0],
[0, 0, 1.0, 0],
[0, 0, 0, 0]])
# 注意:在这里B矩阵中耦合了状态向量x,因此并不是简单的线性模型:
B = np.array([[DT * math.cos(x[2, 0]), 0],
[DT * math.sin(x[2, 0]), 0],
[0.0, DT],
[1.0, 0.0]])
x = F @ x + B @ u
return x
def observation_model(x):
H = np.array([
[1, 0, 0, 0],
[0, 1, 0, 0]
])
z = H @ x
return z
def jacob_f(x, u):
"""
Jacobian of Motion Model
motion model
x_{t+1} = x_t+v*dt*cos(yaw)
y_{t+1} = y_t+v*dt*sin(yaw)
yaw_{t+1} = yaw_t+omega*dt
v_{t+1} = v{t}
so
dx/dyaw = -v*dt*sin(yaw)
dx/dv = dt*cos(yaw)
dy/dyaw = v*dt*cos(yaw)
dy/dv = dt*sin(yaw)
"""
yaw = x[2, 0]
v = u[0, 0]
jF = np.array([
[1.0, 0.0, -DT * v * math.sin(yaw), DT * math.cos(yaw)],
[0.0, 1.0, DT * v * math.cos(yaw), DT * math.sin(yaw)],
[0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 0.0, 1.0]])
return jF
def jacob_h():
# Jacobian of Observation Model
jH = np.array([
[1, 0, 0, 0],
[0, 1, 0, 0]
])
return jH
def ekf_estimation(xEst, PEst, z, u):
# Predict
xPred = motion_model(xEst, u)
jF = jacob_f(xEst, u)
PPred = jF @ PEst @ jF.T + Q
# Update
jH = jacob_h()
zPred = observation_model(xPred)
y = z - zPred
S = jH @ PPred @ jH.T + R
K = PPred @ jH.T @ np.linalg.inv(S)
xEst = xPred + K @ y
PEst = (np.eye(len(xEst)) - K @ jH) @ PPred
return xEst, PEst
def plot_covariance_ellipse(xEst, PEst): # pragma: no cover
Pxy = PEst[0:2, 0:2]
eigval, eigvec = np.linalg.eig(Pxy)
if eigval[0] >= eigval[1]:
bigind = 0
smallind = 1
else:
bigind = 1
smallind = 0
t = np.arange(0, 2 * math.pi + 0.1, 0.1)
a = math.sqrt(eigval[bigind])
b = math.sqrt(eigval[smallind])
x = [a * math.cos(it) for it in t]
y = [b * math.sin(it) for it in t]
angle = math.atan2(eigvec[bigind, 1], eigvec[bigind, 0])
rot = np.array([[math.cos(angle), math.sin(angle)],
[-math.sin(angle), math.cos(angle)]])
fx = rot @ (np.array([x, y]))
px = np.array(fx[0, :] + xEst[0, 0]).flatten()
py = np.array(fx[1, :] + xEst[1, 0]).flatten()
plt.plot(px, py, "--r")
def main():
print(__file__ + " start!!")
time = 0.0
# State Vector [x y yaw v]'
xEst = np.zeros((4, 1)) # 初始值全部为0
xTrue = np.zeros((4, 1))
PEst = np.eye(4) #用一个对角都是1的矩阵表示状态协方差矩阵初始值
xDR = np.zeros((4, 1)) # Dead reckoning
# history
hxEst = xEst
hxTrue = xTrue
hxDR = xTrue
hz = np.zeros((2, 1))
while SIM_TIME >= time:
time += DT
u = calc_input()
xTrue, z, xDR, ud = observation(xTrue, xDR, u)
xEst, PEst = ekf_estimation(xEst, PEst, z, ud)
# store data history
hxEst = np.hstack((hxEst, xEst))
hxDR = np.hstack((hxDR, xDR))
hxTrue = np.hstack((hxTrue, xTrue))
hz = np.hstack((hz, z))
if show_animation:
plt.cla()
# for stopping simulation with the esc key.
plt.gcf().canvas.mpl_connect('key_release_event',
lambda event: [exit(0) if event.key == 'escape' else None])
plt.plot(hz[0, :], hz[1, :], ".g")
plt.plot(hxTrue[0, :].flatten(),
hxTrue[1, :].flatten(), "-b")
plt.plot(hxDR[0, :].flatten(),
hxDR[1, :].flatten(), "-k")
plt.plot(hxEst[0, :].flatten(),
hxEst[1, :].flatten(), "-r")
plot_covariance_ellipse(xEst, PEst)
plt.axis("equal")
plt.grid(True)
plt.pause(0.001)
if __name__ == '__main__':
main()