aaltovision / deep-speed-constrained-ins

Codes for Deep Learning Based Speed Estimation for Constraining Strapdown Inertial Navigation on Smartphones

Deep Learning Based Speed Estimation for Constraining Strapdown Inertial Navigation on Smartphones

Santiago Cortés · Arno Solin · Juho Kannala

Summary

Strapdown inertial navigation systems are sensitive to the quality of the data provided by the accelerometer and gyroscope. Low-grade IMUs in handheld smart-devices pose a problem for inertial odometry on these devices. We propose a scheme for constraining the inertial odometry problem by complementing non-linear state estimation by a CNN-based deep-learning model for inferring the momentary speed based on a window of IMU samples.

This repository provides the codes for replicationg the speed regression setup in [1]. Please, if you use this code/data, please cite the original paper presenting it.

Dependencies:

Ubuntu 16.04 and python 2.7 (including numpy and matplotlib) were used in all the tests.

The following Python packages were also used

• pytorch (0.3)
• pandas (0.23)

Download and prepare training data

Download and unzip ADVIO dataset files (see ADVIO).

cd data
for i in $(seq -f "%02g" 1 23);
do
  wget -O advio-$i.zip https://zenodo.org/record/1321157/files/advio-$i.zip
  unzip advio-$i.zip
  rm advio-$i.zip	
done
cd ..

Synchronize the accelerometer and gyroscope in the ADVIO data.

cd python
python sync-data.py

Produce results

cd python
python DCI-training-0.0.2.py

References

[1] Santiago Cortés, Arno Solin, and Juho Kannala, “Deep Learning Based Speed Estimation for Constraining Strapdown Inertial Navigation on Smartphones”, IEEE International Workshop on Machine Learning for Signal Processing (MLSP), Aalborg, Denmark, 2018. [arXiv]

License

This software is distributed under the GNU General Public License (version 3 or later); please refer to the file LICENSE.txt, included with the software, for details.

deep-speed-constrained-ins's Issues

what the mean of column lthe labels csv ?

0,0,0,0.05
0,0,1,0.48
0,0,2,1.04
0,0,1,2.05
0,0,2,2.22
0,0,0,-1
0,1,0,-2
0,2,0,0.05
0,2,1,0.49
0,2,0,-1
0,3,0,0.03
0,3,1,1.1
0,3,2,1.25
0,3,1,1.35
0,3,2,1.51
0,3,1,-1
0,4,0,0.03
0,4,1,1.52
0,4,2,2.08
0,4,1,2.18
0,4,2,2.31
0,4,0,-1
1,0,0,0.17
1,0,3,0.46
1,0,1,1.21
1,0,3,1.43
1,0,1,2.18
1,0,3,2.38
1,0,1,3.16
1,0,3,3.36
1,0,1,4.13
1,0,0,-1
1,1,0,0.07
1,1,1,-1
1,2,0,0.05
1,2,1,-1
1,3,0,-2
1,4,0,0.15
1,4,2,0.35
1,4,1,-1
1,5,0,0.08
1,5,1,-1
1,6,0,-2
1,7,0,0.13
1,7,1,0.35
1,7,3,0.5
1,7,1,1.22
1,7,3,1.4
1,7,0,-1
1,8,0,0.06
1,8,1,0.28
1,8,3,0.45
1,8,1,1.11
1,8,3,1.28
1,8,0,-1
1,9,0,0.05
1,9,1,-1
1,10,1,-1
1,11,1,-1
2,0,0,0.04
2,0,1,0.26
2,0,0,1.03
2,0,4,1.23
2,0,1,1.51
2,0,2,2.1
2,0,1,2.31
2,0,2,2.46
2,0,1,-1
2,1,1,0.42
2,1,2,0.57
2,1,1,1.58
2,1,2,-1
3,0,1,-1
3,1,1,-1
4,0,1,-1
4,0,1,-2
5,0,0,-1
5,1,0,-1
5,2,0,-1
6,0,1,-1

Could not reproduce pretrained model

I have followed the instructions for downloading and preparing training data and tried to train a model by changing load_model to False and both save_model and train_model to True in DCI-training-0.0.2.py. I also needed to change two lines because I am using pytorch 1.0.0.

I have changed this line:
acc_loss +=np.sum(loss.data[0])

to this:
acc_loss +=np.sum(loss.item())

And I have changed this line:
val_loss +=np.sum(loss.data[0])

to this:
val_loss +=np.sum(loss.item())

However, I am getting the following results:

The final values for loss and validation loss were 2.2964887819458952 and 2.3019133200951383, respectively. Is this expected? Should I change any other parameter (like number of epochs) in order to reproduce the pretrained model?

I am using Ubuntu 18.04.1 LTS, Python 2.7.15rc1 and pandas 0.23.4.

Best regards.