In [31]:
import numpy as np
import matplotlib.dates as mdates
import matplotlib.pyplot as plt
from scipy.stats import norm
from sympy import Symbol, symbols, Matrix, sin, cos
from sympy.interactive import printing
printing.init_printing()

In [32]:
%matplotlib inline
fw = 10 # figure width


# Extended Kalman Filter Implementation for Constant Turn Rate and Velocity (CTRV) Vehicle Model in Python¶

Wikipedia writes: In the extended Kalman filter, the state transition and observation models need not be linear functions of the state but may instead be differentiable functions.

$\boldsymbol{x}_{k} = g(\boldsymbol{x}_{k-1}, \boldsymbol{u}_{k-1}) + \boldsymbol{w}_{k-1}$

$\boldsymbol{z}_{k} = h(\boldsymbol{x}_{k}) + \boldsymbol{v}_{k}$

Where $w_k$ and $v_k$ are the process and observation noises which are both assumed to be zero mean Multivariate Gaussian noises with covariance matrix $Q$ and $R$ respectively.

The function $g$ can be used to compute the predicted state from the previous estimate and similarly the function $h$ can be used to compute the predicted measurement from the predicted state. However, $g$ and $h$ cannot be applied to the covariance directly. Instead a matrix of partial derivatives (the Jacobian matrix) is computed.

At each time step, the Jacobian is evaluated with current predicted states. These matrices can be used in the Kalman filter equations. This process essentially linearizes the non-linear function around the current estimate.

Situation covered: You have a velocity sensor, which measures the vehicle speed ($v$) in heading direction ($\psi$) and a yaw rate sensor ($\dot \psi$) which both have to fused with the position ($x$ & $y$) from a GPS sensor.

## State Vector - Constant Turn Rate and Velocity Vehicle Model (CTRV)¶

Constant Turn Rate, Constant Velocity Model for a vehicle

$$x_k= \left[ \matrix{ x \\ y \\ \psi \\ v \\ \dot\psi} \right] = \left[ \matrix{ \text{Position X} \\ \text{Position Y} \\ \text{Heading} \\ \text{Velocity} \\ \text{Yaw Rate}} \right]$$
In [33]:
numstates=5 # States


The Intertial Measurement Unit (IMU) acquired data 50 times a second, the GPS sensor just with 10Hz, so we only have a position correction every 5th Kalman filter cycle.

In [34]:
dt = 1.0/50.0 # Sample Rate of the Measurements is 50Hz
dtGPS=1.0/10.0 # Sample Rate of GPS is 10Hz


All symbolic calculations are made with Sympy. Thanks!

In [35]:
vs, psis, dpsis, dts, xs, ys, lats, lons = symbols('v \psi \dot\psi T x y lat lon')

gs = Matrix([[xs+(vs/dpsis)*(sin(psis+dpsis*dts)-sin(psis))],
[ys+(vs/dpsis)*(-cos(psis+dpsis*dts)+cos(psis))],
[psis+dpsis*dts],
[vs],
[dpsis]])
state = Matrix([xs,ys,psis,vs,dpsis])


## Dynamic Matrix¶

This formulas calculate how the state is evolving from one to the next time step

In [36]:
gs

Out[36]:
$$\left[\begin{matrix}x + \frac{v}{\dot\psi} \left(- \sin{\left (\psi \right )} + \sin{\left (T \dot\psi + \psi \right )}\right)\\y + \frac{v}{\dot\psi} \left(\cos{\left (\psi \right )} - \cos{\left (T \dot\psi + \psi \right )}\right)\\T \dot\psi + \psi\\v\\\dot\psi\end{matrix}\right]$$

### Calculate the Jacobian of the Dynamic Matrix with respect to the state vector¶

In [37]:
state

Out[37]:
$$\left[\begin{matrix}x\\y\\\psi\\v\\\dot\psi\end{matrix}\right]$$
In [38]:
gs.jacobian(state)

Out[38]:
$$\left[\begin{matrix}1 & 0 & \frac{v}{\dot\psi} \left(- \cos{\left (\psi \right )} + \cos{\left (T \dot\psi + \psi \right )}\right) & \frac{1}{\dot\psi} \left(- \sin{\left (\psi \right )} + \sin{\left (T \dot\psi + \psi \right )}\right) & \frac{T v}{\dot\psi} \cos{\left (T \dot\psi + \psi \right )} - \frac{v}{\dot\psi^{2}} \left(- \sin{\left (\psi \right )} + \sin{\left (T \dot\psi + \psi \right )}\right)\\0 & 1 & \frac{v}{\dot\psi} \left(- \sin{\left (\psi \right )} + \sin{\left (T \dot\psi + \psi \right )}\right) & \frac{1}{\dot\psi} \left(\cos{\left (\psi \right )} - \cos{\left (T \dot\psi + \psi \right )}\right) & \frac{T v}{\dot\psi} \sin{\left (T \dot\psi + \psi \right )} - \frac{v}{\dot\psi^{2}} \left(\cos{\left (\psi \right )} - \cos{\left (T \dot\psi + \psi \right )}\right)\\0 & 0 & 1 & 0 & T\\0 & 0 & 0 & 1 & 0\\0 & 0 & 0 & 0 & 1\end{matrix}\right]$$

It has to be computed on every filter step because it consists of state variables!

To Sympy Team: A .to_python and .to_c and .to_matlab whould be nice to generate code, like it already works with print latex().

## Initial Uncertainty $P_0$¶

Initialized with $0$ means you are pretty sure where the vehicle starts

In [39]:
P = np.diag([1000.0, 1000.0, 1000.0, 1000.0, 1000.0])
print(P, P.shape)

fig = plt.figure(figsize=(5, 5))
im = plt.imshow(P, interpolation="none", cmap=plt.get_cmap('binary'))
plt.title('Initial Covariance Matrix $P$')
ylocs, ylabels = plt.yticks()
# set the locations of the yticks
plt.yticks(np.arange(6))
# set the locations and labels of the yticks
plt.yticks(np.arange(5),('$x$', '$y$', '$\psi$', '$v$', '$\dot \psi$'), fontsize=22)

xlocs, xlabels = plt.xticks()
# set the locations of the yticks
plt.xticks(np.arange(6))
# set the locations and labels of the yticks
plt.xticks(np.arange(5),('$x$', '$y$', '$\psi$', '$v$', '$\dot \psi$'), fontsize=22)

plt.xlim([-0.5,4.5])
plt.ylim([4.5, -0.5])

from mpl_toolkits.axes_grid1 import make_axes_locatable
divider = make_axes_locatable(plt.gca())
plt.colorbar(im, cax=cax)

plt.tight_layout()

(array([[ 1000.,     0.,     0.,     0.,     0.],
[    0.,  1000.,     0.,     0.,     0.],
[    0.,     0.,  1000.,     0.,     0.],
[    0.,     0.,     0.,  1000.,     0.],
[    0.,     0.,     0.,     0.,  1000.]]), (5, 5))


## Process Noise Covariance Matrix Q¶

"The state uncertainty model models the disturbances which excite the linear system. Conceptually, it estimates how bad things can get when the system is run open loop for a given period of time." - Kelly, A. (1994). A 3D state space formulation of a navigation Kalman filter for autonomous vehicles, (May). Retrieved from http://oai.dtic.mil/oai/oai?verb=getRecord&metadataPrefix=html&identifier=ADA282853

In [40]:
sGPS     = 0.5*8.8*dt**2  # assume 8.8m/s2 as maximum acceleration, forcing the vehicle
sCourse  = 0.1*dt # assume 0.1rad/s as maximum turn rate for the vehicle
sVelocity= 8.8*dt # assume 8.8m/s2 as maximum acceleration, forcing the vehicle
sYaw     = 1.0*dt # assume 1.0rad/s2 as the maximum turn rate acceleration for the vehicle

Q = np.diag([sGPS**2, sGPS**2, sCourse**2, sVelocity**2, sYaw**2])
print(Q, Q.shape)

(array([[  3.09760000e-06,   0.00000000e+00,   0.00000000e+00,
0.00000000e+00,   0.00000000e+00],
[  0.00000000e+00,   3.09760000e-06,   0.00000000e+00,
0.00000000e+00,   0.00000000e+00],
[  0.00000000e+00,   0.00000000e+00,   4.00000000e-06,
0.00000000e+00,   0.00000000e+00],
[  0.00000000e+00,   0.00000000e+00,   0.00000000e+00,
3.09760000e-02,   0.00000000e+00],
[  0.00000000e+00,   0.00000000e+00,   0.00000000e+00,
0.00000000e+00,   4.00000000e-04]]), (5, 5))

In [41]:
fig = plt.figure(figsize=(5, 5))
im = plt.imshow(Q, interpolation="none", cmap=plt.get_cmap('binary'))
plt.title('Process Noise Covariance Matrix $Q$')
ylocs, ylabels = plt.yticks()
# set the locations of the yticks
plt.yticks(np.arange(8))
# set the locations and labels of the yticks
plt.yticks(np.arange(7),('$x$', '$y$', '$\psi$', '$v$', '$\dot \psi$'), fontsize=22)

xlocs, xlabels = plt.xticks()
# set the locations of the yticks
plt.xticks(np.arange(8))
# set the locations and labels of the yticks
plt.xticks(np.arange(7),('$x$', '$y$', '$\psi$', '$v$', '$\dot \psi$'), fontsize=22)

plt.xlim([-0.5,4.5])
plt.ylim([4.5, -0.5])

from mpl_toolkits.axes_grid1 import make_axes_locatable
divider = make_axes_locatable(plt.gca())
plt.colorbar(im, cax=cax);


## Real Measurements¶

In [42]:
#path = './../RaspberryPi-CarPC/TinkerDataLogger/DataLogs/2014/'
datafile = '2014-03-26-000-Data.csv'

date, \
time, \
millis, \
ax, \
ay, \
az, \
rollrate, \
pitchrate, \
yawrate, \
roll, \
pitch, \
yaw, \
speed, \
course, \
latitude, \
longitude, \
altitude, \
pdop, \
hdop, \
vdop, \
epe, \
fix, \
satellites_view, \
satellites_used, \
converters={1: mdates.strpdate2num('%H%M%S%f'),
0: mdates.strpdate2num('%y%m%d')},
skiprows=1)

# A course of 0Â° means the Car is traveling north bound
# and 90Â° means it is traveling east bound.
# In the Calculation following, East is Zero and North is 90Â°
# We need an offset.
course =(-course+90.0)

Read '2014-03-26-000-Data.csv' successfully.


## Measurement Function H¶

Matrix $J_H$ is the Jacobian of the Measurement function $h$ with respect to the state. Function $h$ can be used to compute the predicted measurement from the predicted state.

If a GPS measurement is available, the following function maps the state to the measurement.

In [43]:
hs = Matrix([[xs],
[ys],
[vs],
[dpsis]])
hs

Out[43]:
$$\left[\begin{matrix}x\\y\\v\\\dot\psi\end{matrix}\right]$$
In [44]:
JHs=hs.jacobian(state)
JHs

Out[44]:
$$\left[\begin{matrix}1 & 0 & 0 & 0 & 0\\0 & 1 & 0 & 0 & 0\\0 & 0 & 0 & 1 & 0\\0 & 0 & 0 & 0 & 1\end{matrix}\right]$$

If no GPS measurement is available, simply set the corresponding values in $J_h$ to zero.

## Measurement Noise Covariance $R$¶

"In practical use, the uncertainty estimates take on the significance of relative weights of state estimates and measurements. So it is not so much important that uncertainty is absolutely correct as it is that it be relatively consistent across all models" - Kelly, A. (1994). A 3D state space formulation of a navigation Kalman filter for autonomous vehicles, (May). Retrieved from http://oai.dtic.mil/oai/oai?verb=getRecord&metadataPrefix=html&identifier=ADA282853

In [45]:
sGPS = 5.0   # Standard Deviation of GPS Measurement
sspeed = 2.0 # Standard Deviation of the speed measurement
syaw = 0.01  # Standard Deviation of the yawrate measurement
R = np.matrix([[sGPS**2, 0.0, 0.0, 0.0],
[0.0, sGPS**2, 0.0, 0.0],
[0.0, 0.0, sspeed**2, 0.0],
[0.0, 0.0, 0.0, syaw**2]])

print(R, R.shape)

(matrix([[  2.50000000e+01,   0.00000000e+00,   0.00000000e+00,
0.00000000e+00],
[  0.00000000e+00,   2.50000000e+01,   0.00000000e+00,
0.00000000e+00],
[  0.00000000e+00,   0.00000000e+00,   4.00000000e+00,
0.00000000e+00],
[  0.00000000e+00,   0.00000000e+00,   0.00000000e+00,
1.00000000e-04]]), (4, 4))

In [46]:
fig = plt.figure(figsize=(4.5, 4.5))
im = plt.imshow(R, interpolation="none", cmap=plt.get_cmap('binary'))
plt.title('Measurement Noise Covariance Matrix $R$')
ylocs, ylabels = plt.yticks()
# set the locations of the yticks
plt.yticks(np.arange(5))
# set the locations and labels of the yticks
plt.yticks(np.arange(4),('$x$', '$y$', '$v$', '$\dot \psi$'), fontsize=22)

xlocs, xlabels = plt.xticks()
# set the locations of the yticks
plt.xticks(np.arange(5))
# set the locations and labels of the yticks
plt.xticks(np.arange(4),('$x$', '$y$', '$v$', '$\dot \psi$'), fontsize=22)

plt.xlim([-0.5,3.5])
plt.ylim([3.5, -0.5])

from mpl_toolkits.axes_grid1 import make_axes_locatable
divider = make_axes_locatable(plt.gca())
plt.colorbar(im, cax=cax);


## Identity Matrix¶

In [47]:
I = np.eye(numstates)
print(I, I.shape)

(array([[ 1.,  0.,  0.,  0.,  0.],
[ 0.,  1.,  0.,  0.,  0.],
[ 0.,  0.,  1.,  0.,  0.],
[ 0.,  0.,  0.,  1.,  0.],
[ 0.,  0.,  0.,  0.,  1.]]), (5, 5))


## Approx. Lat/Lon to Meters to check Location¶

In [48]:
RadiusEarth = 6378388.0 # m

dx = arc * np.cos(latitude*np.pi/180.0) * np.hstack((0.0, np.diff(longitude))) # in m
dy = arc * np.hstack((0.0, np.diff(latitude))) # in m

mx = np.cumsum(dx)
my = np.cumsum(dy)

ds = np.sqrt(dx**2+dy**2)

GPS=np.hstack((True, (np.diff(ds)>0.0).astype('bool'))) # GPS Trigger for Kalman Filter


## Initial State¶

In [49]:
x = np.matrix([[mx[0], my[0], course[0]/180.0*np.pi, speed[0]/3.6+0.001, yawrate[0]/180.0*np.pi]]).T
print(x, x.shape)

U=float(np.cos(x[2])*x[3])
V=float(np.sin(x[2])*x[3])

plt.quiver(x[0], x[1], U, V)
plt.scatter(float(x[0]), float(x[1]), s=100)
plt.title('Initial Location')
plt.axis('equal')

(matrix([[ 0.        ],
[ 0.        ],
[-4.08756111],
[ 0.67322222],
[-0.32660346]]), (5, 1))

Out[49]:
$$\begin{pmatrix}-0.00015, & 0.00015, & -0.0003, & 0.0003\end{pmatrix}$$

### Put everything together as a measurement vector¶

In [50]:
measurements = np.vstack((mx, my, speed/3.6, yawrate/180.0*np.pi))
# Lenth of the measurement
m = measurements.shape[1]
print(measurements.shape)

(4, 10800)

In [51]:
# Preallocation for Plotting
x0 = []
x1 = []
x2 = []
x3 = []
x4 = []
x5 = []
Zx = []
Zy = []
Px = []
Py = []
Pdx= []
Pdy= []
Pddx=[]
Pddy=[]
Kx = []
Ky = []
Kdx= []
Kdy= []
Kddx=[]
dstate=[]


# Extended Kalman Filter¶

$$x_k= \begin{bmatrix} x \\ y \\ \psi \\ v \\ \dot\psi \end{bmatrix} = \begin{bmatrix} \text{Position X} \\ \text{Position Y} \\ \text{Heading} \\ \text{Velocity} \\ \text{Yaw Rate} \end{bmatrix} = \underbrace{\begin{matrix}x[0] \\ x[1] \\ x[2] \\ x[3] \\ x[4] \end{matrix}}_{\textrm{Python Nomenclature}}$$
In [52]:
for filterstep in range(m):

# Time Update (Prediction)
# ========================
# see "Dynamic Matrix"
if np.abs(yawrate[filterstep])<0.0001: # Driving straight
x[0] = x[0] + x[3]*dt * np.cos(x[2])
x[1] = x[1] + x[3]*dt * np.sin(x[2])
x[2] = x[2]
x[3] = x[3]
x[4] = 0.0000001 # avoid numerical issues in Jacobians
dstate.append(0)
else: # otherwise
x[0] = x[0] + (x[3]/x[4]) * (np.sin(x[4]*dt+x[2]) - np.sin(x[2]))
x[1] = x[1] + (x[3]/x[4]) * (-np.cos(x[4]*dt+x[2])+ np.cos(x[2]))
x[2] = (x[2] + x[4]*dt + np.pi) % (2.0*np.pi) - np.pi
x[3] = x[3]
x[4] = x[4]
dstate.append(1)

# Calculate the Jacobian of the Dynamic Matrix A
# see "Calculate the Jacobian of the Dynamic Matrix with respect to the state vector"
a13 = float((x[3]/x[4]) * (np.cos(x[4]*dt+x[2]) - np.cos(x[2])))
a14 = float((1.0/x[4]) * (np.sin(x[4]*dt+x[2]) - np.sin(x[2])))
a15 = float((dt*x[3]/x[4])*np.cos(x[4]*dt+x[2]) - (x[3]/x[4]**2)*(np.sin(x[4]*dt+x[2]) - np.sin(x[2])))
a23 = float((x[3]/x[4]) * (np.sin(x[4]*dt+x[2]) - np.sin(x[2])))
a24 = float((1.0/x[4]) * (-np.cos(x[4]*dt+x[2]) + np.cos(x[2])))
a25 = float((dt*x[3]/x[4])*np.sin(x[4]*dt+x[2]) - (x[3]/x[4]**2)*(-np.cos(x[4]*dt+x[2]) + np.cos(x[2])))
JA = np.matrix([[1.0, 0.0, a13, a14, a15],
[0.0, 1.0, a23, a24, a25],
[0.0, 0.0, 1.0, 0.0, dt],
[0.0, 0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 1.0]])

# Project the error covariance ahead
P = JA*P*JA.T + Q

# Measurement Update (Correction)
# ===============================
# Measurement Function
hx = np.matrix([[float(x[0])],
[float(x[1])],
[float(x[3])],
[float(x[4])]])

if GPS[filterstep]:
JH = np.matrix([[1.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 1.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 1.0]])
else:
JH = np.matrix([[0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 1.0]])

S = JH*P*JH.T + R
K = (P*JH.T) * np.linalg.inv(S)

# Update the estimate via
Z = measurements[:,filterstep].reshape(JH.shape[0],1)
y = Z - (hx)                         # Innovation or Residual
x = x + (K*y)

# Update the error covariance
P = (I - (K*JH))*P

# Save states for Plotting
x0.append(float(x[0]))
x1.append(float(x[1]))
x2.append(float(x[2]))
x3.append(float(x[3]))
x4.append(float(x[4]))
Zx.append(float(Z[0]))
Zy.append(float(Z[1]))
Px.append(float(P[0,0]))
Py.append(float(P[1,1]))
Pdx.append(float(P[2,2]))
Pdy.append(float(P[3,3]))
Pddx.append(float(P[4,4]))
Kx.append(float(K[0,0]))
Ky.append(float(K[1,0]))
Kdx.append(float(K[2,0]))
Kdy.append(float(K[3,0]))
Kddx.append(float(K[4,0]))


## Plots¶

### Uncertainties¶

In [53]:
fig = plt.figure(figsize=(fw,9))
plt.semilogy(range(m),Px, label='$x$')
plt.step(range(m),Py, label='$y$')
plt.step(range(m),Pdx, label='$\psi$')
plt.step(range(m),Pdy, label='$v$')
plt.step(range(m),Pddx, label='$\dot \psi$')

plt.xlabel('Filter Step')
plt.ylabel('')
plt.title('Uncertainty (Elements from Matrix $P$)')
plt.legend(loc='best',prop={'size':22})

Out[53]:
<matplotlib.legend.Legend at 0x1082154d0>
In [54]:
fig = plt.figure(figsize=(6, 6))
im = plt.imshow(P, interpolation="none", cmap=plt.get_cmap('binary'))
plt.title('Covariance Matrix $P$ (after %i Filter Steps)' % (m))
ylocs, ylabels = plt.yticks()
# set the locations of the yticks
plt.yticks(np.arange(6))
# set the locations and labels of the yticks
plt.yticks(np.arange(5),('$x$', '$y$', '$\psi$', '$v$', '$\dot \psi$'), fontsize=22)

xlocs, xlabels = plt.xticks()
# set the locations of the yticks
plt.xticks(np.arange(6))
# set the locations and labels of the yticks
plt.xticks(np.arange(5),('$x$', '$y$', '$\psi$', '$v$', '$\dot \psi$'), fontsize=22)

plt.xlim([-0.5,4.5])
plt.ylim([4.5, -0.5])

from mpl_toolkits.axes_grid1 import make_axes_locatable
divider = make_axes_locatable(plt.gca())
plt.colorbar(im, cax=cax)

plt.tight_layout()


### Kalman Gains¶

In [55]:
fig = plt.figure(figsize=(fw,9))
plt.step(range(len(measurements[0])),Kx, label='$x$')
plt.step(range(len(measurements[0])),Ky, label='$y$')
plt.step(range(len(measurements[0])),Kdx, label='$\psi$')
plt.step(range(len(measurements[0])),Kdy, label='$v$')
plt.step(range(len(measurements[0])),Kddx, label='$\dot \psi$')

plt.xlabel('Filter Step')
plt.ylabel('')
plt.title('Kalman Gain (the lower, the more the measurement fullfill the prediction)')
plt.legend(prop={'size':18})
plt.ylim([-0.1,0.1]);


## State Vector¶

In [56]:
fig = plt.figure(figsize=(fw,16))

plt.subplot(411)
plt.step(range(len(measurements[0])),x0-mx[0], label='$x$')
plt.step(range(len(measurements[0])),x1-my[0], label='$y$')

plt.title('Extended Kalman Filter State Estimates (State Vector $x$)')
plt.legend(loc='best',prop={'size':22})
plt.ylabel('Position (relative to start) [m]')

plt.subplot(412)
plt.step(range(len(measurements[0])),x2, label='$\psi$')
plt.step(range(len(measurements[0])),(course/180.0*np.pi+np.pi)%(2.0*np.pi) - np.pi, label='$\psi$ (from GPS as reference)')
plt.ylabel('Course')
plt.legend(loc='best',prop={'size':16})

plt.subplot(413)
plt.step(range(len(measurements[0])),x3, label='$v$')
plt.step(range(len(measurements[0])),speed/3.6, label='$v$ (from GPS as reference)')
plt.ylabel('Velocity')
plt.ylim([0, 30])
plt.legend(loc='best',prop={'size':16})

plt.subplot(414)
plt.step(range(len(measurements[0])),x4, label='$\dot \psi$')
plt.step(range(len(measurements[0])),yawrate/180.0*np.pi, label='$\dot \psi$ (from IMU as reference)')
plt.ylabel('Yaw Rate')
plt.ylim([-0.6, 0.6])
plt.legend(loc='best',prop={'size':16})
plt.xlabel('Filter Step')

plt.savefig('Extended-Kalman-Filter-CTRV-State-Estimates.png', dpi=72, transparent=True, bbox_inches='tight')


## Position x/y¶

In [57]:
#%pylab --no-import-all

In [58]:
fig = plt.figure(figsize=(16,9))

# EKF State
#plt.quiver(x0,x1,np.cos(x2), np.sin(x2), color='#94C600', units='xy', width=0.05, scale=0.5)
plt.plot(x0,x1, label='EKF Position')

# Measurements
plt.scatter(mx[::5],my[::5], s=50, label='GPS Measurements')
#cbar=plt.colorbar(ticks=np.arange(20))
#cbar.ax.set_ylabel(u'EPE', rotation=270)
#cbar.ax.set_xlabel(u'm')

# Start/Goal
plt.scatter(x0[0],x1[0], s=60, label='Start', c='g')
plt.scatter(x0[-1],x1[-1], s=60, label='Goal', c='r')

plt.xlabel('X [m]')
plt.ylabel('Y [m]')
plt.title('Position')
plt.legend(loc='best')
plt.axis('equal')
#plt.tight_layout()

#plt.savefig('Extended-Kalman-Filter-CTRV-Position.png', dpi=72, transparent=True, bbox_inches='tight')

Out[58]:
$$\begin{pmatrix}-100.0, & 700.0, & -50.0, & 400.0\end{pmatrix}$$

### Detailed View¶

In [59]:
fig = plt.figure(figsize=(9,4))

# EKF State
#plt.quiver(x0,x1,np.cos(x2), np.sin(x2), color='#94C600', units='xy', width=0.01, scale=0.2, label='Driving Direction')
plt.plot(x0,x1, label='EKF Position', linewidth=5, alpha=0.8)

# Measurements
plt.scatter(mx[::5],my[::5], s=50, label='GPS Measurements')
#cbar=plt.colorbar(ticks=np.arange(20))
#cbar.ax.set_ylabel(u'EPE', rotation=270)
#cbar.ax.set_xlabel(u'm')

plt.xlabel('X [m]')
plt.xlim(80, 120)
plt.ylabel('Y [m]')
plt.ylim(160, 180)
plt.title('Position')
plt.legend(loc='best');
plt.savefig('EKF-Position.png', dpi=150)


# Conclusion¶

As you can see, complicated analytic calculation of the Jacobian Matrices, but it works pretty well.