I have some data that represents the location of an object measured from two different sensors. So, I need to do sensor fusion. The more difficult issue is that the data from each sensor, arrives at essentially a random time. I would like to use pykalman so fuse and smooth the data. How can pykalman handle variable timestamp data?
A simplified sample of the data will look like this:
import pandas as pd
data={'time':\
['10:00:00.0','10:00:01.0','10:00:05.2','10:00:07.5','10:00:07.5','10:00:12.0','10:00:12.5']\
,'X':[10,10.1,20.2,25.0,25.1,35.1,35.0],'Y':[20,20.2,41,45,47,75.0,77.2],\
'Sensor':[1,2,1,1,2,1,2]}
df=pd.DataFrame(data,columns=['time','X','Y','Sensor'])
df.time=pd.to_datetime(df.time)
df=df.set_index('time')
And this:
df
Out[130]:
X Y Sensor
time
2017-12-01 10:00:00.000 10.0 20.0 1
2017-12-01 10:00:01.000 10.1 20.2 2
2017-12-01 10:00:05.200 20.2 41.0 1
2017-12-01 10:00:07.500 25.0 45.0 1
2017-12-01 10:00:07.500 25.1 47.0 2
2017-12-01 10:00:12.000 35.1 75.0 1
2017-12-01 10:00:12.500 35.0 77.2 2
For the sensor fusing issue, I think that I can just reshape the data so that I have positions X1,Y1,X2,Y2 with a bunch of missing values, instead of just X,Y. (This was related: https://stackoverflow.com/questions/47386426/2-sensor-readings-fusion-yaw-pitch )
So then my data can look like this:
df['X1']=df.X[df.Sensor==1]
df['Y1']=df.Y[df.Sensor==1]
df['X2']=df.X[df.Sensor==2]
df['Y2']=df.Y[df.Sensor==2]
df
Out[132]:
X Y Sensor X1 Y1 X2 Y2
time
2017-12-01 10:00:00.000 10.0 20.0 1 10.0 20.0 NaN NaN
2017-12-01 10:00:01.000 10.1 20.2 2 NaN NaN 10.1 20.2
2017-12-01 10:00:05.200 20.2 41.0 1 20.2 41.0 NaN NaN
2017-12-01 10:00:07.500 25.0 45.0 1 25.0 45.0 25.1 47.0
2017-12-01 10:00:07.500 25.1 47.0 2 25.0 45.0 25.1 47.0
2017-12-01 10:00:12.000 35.1 75.0 1 35.1 75.0 NaN NaN
2017-12-01 10:00:12.500 35.0 77.2 2 NaN NaN 35.0 77.2
The docs for pykalman indicate that it can handle missing data, but is that correct?
But, the docs for pykalman are not at all clear about the variable time issue. The doc just says:
"Both the Kalman Filter and Kalman Smoother are able to use parameters which vary with time. In order to use this, one need only pass in an array n_timesteps in length along its first axis:"
>>> transition_offsets = [[-1], [0], [1], [2]]
>>> kf = KalmanFilter(transition_offsets=transition_offsets, n_dim_obs=1)
I have not been able to find any examples of using a pykalman Smoother with variable time steps. So, any guidance, examples or even an example using my above data, would be very helpful. I is not necessary to use pykalman, but it seems like a useful tool to smooth this data.
*****Additional code added below @Anton I made a version of your helpful code that uses the smooth function. The strange thing is that it seems to treat every observation with the same weight and has the trajectory go through every single one. Even, if I have a large different between the sensor variance values. I would expect that around the 5.4,5.0 point, the filtered trajectory should go closer to the Sensor 1 point, since that one has a lower variance. Instead the trajectory goes exactly to each point, and makes a big turn to get there.
from pykalman import KalmanFilter
import numpy as np
import matplotlib.pyplot as plt
# reading data (quick and dirty)
Time=[]
RefX=[]
RefY=[]
Sensor=[]
X=[]
Y=[]
for line in open('data/dataset_01.csv'):
f1, f2, f3, f4, f5, f6 = line.split(';')
Time.append(float(f1))
RefX.append(float(f2))
RefY.append(float(f3))
Sensor.append(float(f4))
X.append(float(f5))
Y.append(float(f6))
# Sensor 1 has a higher precision (max error = 0.1 m)
# Sensor 2 has a lower precision (max error = 0.3 m)
# Variance definition through 3-Sigma rule
Sensor_1_Variance = (0.1/3)**2;
Sensor_2_Variance = (0.3/3)**2;
# Filter Configuration
# time step
dt = Time[2] - Time[1]
# transition_matrix
F = [[1, 0, dt, 0],
[0, 1, 0, dt],
[0, 0, 1, 0],
[0, 0, 0, 1]]
# observation_matrix
H = [[1, 0, 0, 0],
[0, 1, 0, 0]]
# transition_covariance
Q = [[1e-4, 0, 0, 0],
[ 0, 1e-4, 0, 0],
[ 0, 0, 1e-4, 0],
[ 0, 0, 0, 1e-4]]
# observation_covariance
R_1 = [[Sensor_1_Variance, 0],
[0, Sensor_1_Variance]]
R_2 = [[Sensor_2_Variance, 0],
[0, Sensor_2_Variance]]
# initial_state_mean
X0 = [0,
0,
0,
0]
# initial_state_covariance - assumed a bigger uncertainty in initial velocity
P0 = [[ 0, 0, 0, 0],
[ 0, 0, 0, 0],
[ 0, 0, 1, 0],
[ 0, 0, 0, 1]]
n_timesteps = len(Time)
n_dim_state = 4
filtered_state_means = np.zeros((n_timesteps, n_dim_state))
filtered_state_covariances = np.zeros((n_timesteps, n_dim_state, n_dim_state))
import numpy.ma as ma
obs_cov=np.zeros([n_timesteps,2,2])
obs=np.zeros([n_timesteps,2])
for t in range(n_timesteps):
if Sensor[t] == 0:
obs[t]=None
else:
obs[t] = [X[t], Y[t]]
if Sensor[t] == 1:
obs_cov[t] = np.asarray(R_1)
else:
obs_cov[t] = np.asarray(R_2)
ma_obs=ma.masked_invalid(obs)
ma_obs_cov=ma.masked_invalid(obs_cov)
# Kalman-Filter initialization
kf = KalmanFilter(transition_matrices = F,
observation_matrices = H,
transition_covariance = Q,
observation_covariance = ma_obs_cov, # the covariance will be adapted depending on Sensor_ID
initial_state_mean = X0,
initial_state_covariance = P0)
filtered_state_means, filtered_state_covariances=kf.smooth(ma_obs)
# extracting the Sensor update points for the plot
Sensor_1_update_index = [i for i, x in enumerate(Sensor) if x == 1]
Sensor_2_update_index = [i for i, x in enumerate(Sensor) if x == 2]
Sensor_1_update_X = [ X[i] for i in Sensor_1_update_index ]
Sensor_1_update_Y = [ Y[i] for i in Sensor_1_update_index ]
Sensor_2_update_X = [ X[i] for i in Sensor_2_update_index ]
Sensor_2_update_Y = [ Y[i] for i in Sensor_2_update_index ]
# plot of the resulted trajectory
plt.plot(RefX, RefY, "k-", label="Real Trajectory")
plt.plot(Sensor_1_update_X, Sensor_1_update_Y, "ro", label="Sensor 1")
plt.plot(Sensor_2_update_X, Sensor_2_update_Y, "bo", label="Sensor 2")
plt.plot(filtered_state_means[:, 0], filtered_state_means[:, 1], "g.", label="Filtered Trajectory", markersize=1)
plt.grid()
plt.legend(loc="upper left")
plt.show()
For a Kalman filter it is useful to represent the input data with a constant time step. Your sensors send data randomly, so you can define the smallest significant time step for your system and discretize the time axis with this step.
For example one of your sensors sends data approximately each 0.2 seconds and the second one each 0.5 seconds. So the smallest time step could be 0.01 seconds (here you need to find a trade-off between computational time and desired precision).
Your data would look like this:
Now you need to call the Pykalman function filter_update depending on your observations. If there is no observation, the filter predicts the next state based on the previous one. If there is an observation, it corrects the system state.
Probably your sensors have different accuracy. So you can specify the observation covariance depending on the sensor variance.
To demonstrate the idea I generated a 2D-trajectory and randomly put measurements of 2 sensors with different accuracy.
Here is the result:
I chose really bad parameters on purpose, so one can see the prediction and correction steps. Between the sensor updates the filter predicts the trajectory based on the constant velocity from the previous step. As soon as an update comes the filter corrects the position according to the sensor's variance. The precision of the second sensor is very bad, so it influences the system with a lower weight.
Here is my python code:
I put the csv-file here so you can execute the code.
I hope I could help you.
UPDATE
Some information to your suggestion about a variable transition matrix. I would say it depends on the availability of your sensors and on the requirements to the estimation result.
Here I plotted the same estimation both with a constant and a variable transition matrix (I changed the transition covariance matrix, otherwise the estimation was too bad because of the high filter "stiffness"):
As you can see the estimated position of the yellow markers is pretty good. BUT! you have no information between the sensor readings. Using a variable transition matrix you avoid the prediction step between readings and have no idea what happens to the system. It can be good enough if your readings come with a high rate, but otherwise it can be a disadvantage.
Here is the updated code:
Another important point that I did not implement in this code: while using a variable transition matrix you need to vary the transition covariance matrix as well (depending on the current dt).
It's an interesting topic. Let me know what kind of estimation fits to your problem best.