Waveform-Guide Transformation of IMU Measurements for Smartphone-Based Localization

The raw data are consist of three pattern with below types : hodling(chest), swing, pocket

 - type 1 : repeat 10 step walking & turn left 360 degree

 - type 2 : repeat 10 step walking & turn left 180 degree

 - type 3 : repeat 10 step walking & turn left 275 degree

 - type 4 : repeat 10 step walking & turn left 90 degree

 - type 5 : repeat 10 step walking & turn left 315 degree

 - type 6 : repeat 10 step walking & turn left 225 degree

 - type 7 : repeat 10 step walking & turn left 135 degree

 - type 8 : repeat 10 step walking & turn left 45 degree

 - type 12 : repeat 10 step walking & turn left 180 and 360 degrees.

 

 - type 11 : repeat 10 step walking & turn right 360 degree

 - type 22 : repeat 10 step walking & turn right 180 degree

 - type 33 : repeat 10 step walking & turn right 275 degree

 - type 44 : repeat 10 step walking & turn right 90 degree

 - type 55 : repeat 10 step walking & turn right 315 degree

 - type 66 : repeat 10 step walking & turn right 225 degree

 - type 77 : repeat 10 step walking & turn right 135 degree

 - type 88 : repeat 10 step walking & turn right 45 degree

 

-------------------------------------------------------------------------------------------------------

The raw data progresses to the time-aligned process (due to the accelerometer and gyroscope unsynced problem) and to the high pass & low pass filter.

The test and train data consist of below categories: 

 

{'time';'Accm_x';'Accm_y';'Accm_z';'Accm_mag'; 'Gyrm_x';'Gyrm_y';'Gyrm_z';'Gyrm_mag';...

'Accm_Out'; 'Gyrm_Out'; 'heading_change'; 'length_diff';'theta';...

'Accm_x_raw'; 'Accm_y_raw'; 'Accm_z_raw'; 'Gyrm_x_raw'; 'Gyrm_y_raw'; 'Gyrm_z_raw'};

 

'time' is a measurement of time.

'Accm_x', 'Accm_y', 'Accm_z' 'Gyrm_x' 'Gyrm_y' 'Gyrm_z' are the pre-processed data described in [1].

'Accm_mag' and 'Gyrm_mag' are the magnitude of the accelerometer and gyroscope measurements, e.g., Accm_mag = norm(Accm_x_raw, Accm_y_raw, Accm_z_raw).

'Accm_Out' and 'Gyrm_Out' are reference waveforms (ground-truth).

'heading_change' is a heading difference between times.

'length_diff' is a length difference between times.

'theta' is a heading direction estimated in a holding pattern through [2].

'Accm_x,y,z_raw' and 'Gyrm_x,y,z_raw' are raw measurements with time synced.

 

-------------------------------------------------------------------------------------------------------

The train data set is 

 - type 2: path 1,2 #2

 - type 4: path 1:3, 14:19 #9 

 - type 7: path 1:3, 5:12 #11

 - type 8: path 2,3, 5:13 #11

 - type 12: path 1:5, 7,8, 12 #8

 - type 22: path 1:9, 11 #10

 - type 44: 5,6, 8:12 #6

 - type 77: path 1:12 #12

 - type 88: path 1:15 #15

The total # is 84.

 

The test data set is

 - type 2: path 4 #1

 - type 4: path 4, 20,21 #3

 - type 7: path 13:16 #4

 - type 8: path 4, 14:16 #4

 - type 22: path 10, 12:14 #4

 - type 44: 7, 13 #2

 - type 77: path 13:16 #4

 - type 88: path 17, 19:21, 23 #5

The total # is 27.

-------------------------------------------------------------------------------------------------------

Conversion to raw data to train and test data

 

1) Open matlab file 'DataAcquisition_Kinetic_v2_1_full_raw' in raw_data

2) Set a parameter and File path

 - Train stride 2, Test stride 6.

 - isTest = true for test data generation.

 - isCategory = true for train data generation.

 - window size is LPF paramter for the acceleromter and gyroscope. default: [2, 6]

 - isRRM = true for train data generation.

3) Set a path and type for data generation.

 

-------------------------------------------------------------------------------------------------------

The deep learning model consists of the following structure.

 1. 1d_resnet accm (Proposed model 1)

 - input (Accm_mag, Gyrm_mag): 2x80

 - output (1d_resnet_accm_out): 1x80

 - loss: MSE (Accm_out, 1d_resnet accm)

 

2. 1d_resnet gyrm (Proposed model 2)

 - input (Accm_x,y,z,mag, Gyrm_x,y,z,mag, 1d_resnet accm): 9x80

 - output (1d_resnet_gyrm_out): 1x80

 - loss: MSE (Gyrm_out, 1d_resnet_gyrm_out): 1x80

 

3. PDRnet [3] (Comparision model)

 - input (Accm_x,y,z_raw, Gyrm_x,y,z_raw): 6x80

 - output (heading_change, length_diff): 2x80

-------------------------------------------------------------------------------------------------------

Reference

[1] Kyuwon Han, Seung Min Yu, Seung-Woo Ko, and Seong-Lyun Kim, "Waveform-Guide Transformation of IMU Measurements for Smartphone-Based Localization," has been submitted to the IEEE for for possible publication.

[2] W. Kang and Y. Han, "SmartPDR: Smartphone-based pedestrian dead reckoning for indoor localization," IEEE Sensors Journal, vol. 15, no. 5, pp. 2906-2916.

[3] O. Asraf, F. Shama and I. Klein, "PDRNet: A Deep-Learning Pedestrian Dead Reckoning Framework," in IEEE Sensors Journal, vol. 22, no. 6, pp. 4932-4939, 15 March15, 2022, doi: 10.1109/JSEN.2021.3066840.