This research presents the use of deep learning based ensemble classi er to perform indoor localization with heterogeneous devices. Features extracted from magnetic data of Galaxy S8 are fed into neural networks (NNs) for training. The experiments are performed with S8 and G6 smartphones to nd out the impact of device dependence on localization accuracy. Results demonstrate that NNs can play a potential role for precise indoor localization. The proposed approach is able to achieve a localization accuracy of 2.5 m at 50% on two di erent devices. Mean error for S8 and G6 is 2.61 m and 2.95 m, respectively.
strength varies from 25 Tesla to 65 Tesla over the globe. On the other hand, man made
construction obstruct the magnetic eld and alter it to cause anomalies. Such magnetic anomalies
are observed to exhibit unique behavior and used as ngerprint in many research works [
Deep learning has recently been utilized to solve many problems and indoor localization is no exception. DNN and convolution neural networks (CNN) have been used for indoor scene recognition, object detection, etc. We make the use of ensemble learning wherein more than one DNN are trained and each DNN serves as a location classi er. The prediction from each of these classi ers is then employed to nd the nal location of the user. Deep learning is a data intensive technique and requires a large amount of data for training. We have collected thousands of magnetic samples for this purpose. The key contribution of this research is the proposal of a smartphone sensors base indoor localization approach which works with multiple neural networks (NN) to predict user's current location. The proposed approach is tested with heterogeneous devices including Galaxy S8 and LG G6 to evaluate the localization accuracy.
The rest of the paper is organized in the following manner. Section 2 overviews few works related to this research. Section 3 describes the proposed approach while Section 4 details the experiment setup and analyzes the results. Finally, conclusion is given in Section 5. 2
The application of magnetic eld data for indoor localization has been investigated by many
research works . Such research include the analysis of properties of magnetic eld data that can be
used for localization, as well as, the impact of various devices usage, and the orientations of these
devices [
An indoor localization system is presented in [
The above mentioned research works are limited by two factors in essence. First problem is the use of Wi-Fi signals which are vulnerable to dynamic factors. Secondly the impact of device heterogeneity is not studied very well. Additionally the use of smartphone camera consumes the battery very fast and is not an e cient solution. It is noteworthy to point out that deep learning has been utilized on smartphone camera images alone. We aim to use deep learning on the magnetic eld data to perform indoor localization. 3
This section provides the details of the proposed approach. The rst task is to features which are fed into the NNs. nd suitable 3.1
Features Selection The major limitation of using the magnetic eld is the device dependence. The intensity of collected magnetic data may be di erent depending on the sensitivity of the installed magnetometer in various smartphones. Another shortcoming of magnetic data is its low dimensionality. We can use only magnetic x, y, and z data as a ngerprint. These values may be very similar at multiple locations, especially in large space. So, contrary to using the magnetic eld data, we aim to work with important features of this data.
We initially shortlisted a total of 18 features including 'minimum', 'maximum', 'mean', 'trimmed mean', 'median', 'root mean square', 'standard deviation', 'interquartile', 'percentile (1, 50, 75, 99)', 'mean absolute deviation', 'coe cient of variance', 'kurtosis ', 'Shanon's entropy ', and 'skewness ' for this purpose. Features including 'coe cient of variance', 'kurtosis ', 'Shanon's entropy ', and 'skewness ' are dropped due to their little correlation to the classi cation label. The correlation of the features is shown in Figure 1. The architecture of the proposed approach is shown in Figure 2. The features extracted from the magnetic data are standardized and fed into neural networks. We train three di erent NN to make the ensemble.
Every NN has di erent number of layers, as well as, the containing neurons. Similarly, the internal structure of fully connected layers is di erent. The structure of NNs is shown in Figure 3. During the localization phase, the features extracted from user collected magnetic data are used by trained NNs to predict user current location. For this purpose, we use three consecutive frames of 2 sec each which are considered as T1, T2, and T3. For T1, the predictions from three NNs are taken to form the location candidates (Lc).
Lc = 9(PNN1T 1 [ PNN2T 1 ) [ (PNN1T 1 [ PNN3T 1 ) [ (PNN2T 1 [ PNN3T 1 ) (1) where P shows the prediction made by NN and 9 shows that unique predictions are considered alone. We take top 10 predicted locations from each NN to formulate Lc. NNs predictions at T2 and T3 help to re ne Lc. So, if the predictions for T2 are in the area as shown in Figure 4, they are added in Lc for T3, else, they are discarded. The encircled area is estimated user traveled distance at a medium speed.
Algorithm 1 Find user location 1: Lc
f indLocCandidates(PNN1T1 ; PNN2T1 ; PNN3T1 ) 2: for T 2 to 3 do 3: Lc ref ineLocCandidates(Lc; PNN1T ; PNN2T ; PNN3T ); 4: (xT ; yT ) calApproxP os(SlT ; T ); 5: end for
calCenteroid(Lc);
With each T , Lc are updated as user moves to another location. We use accelerometer and
gyroscope data for this purpose. Step detection is performed using the algorithm proposed in [
Wemigobdeerlg model [
p Sl = k 4 amax amin where amax, and amin shows the maximum and minimum acceleration.
Softmax
Softmax
Loss Val. loss Accuracy Val. accuracy Loss Val. loss Accuracy Val. accuracy
Loss Val. loss Accuracy Val. accuracy (2) (3) (4) ReLU1
ReLU2
ReLU3 AdReaLUm4 nnReLUm5odReeLUl6
ReLU7
ReLU8
ReLU9 F.connected1 F.connected2 F.connected3 F.connected4 F.connected5 F.connected6 F.connected7 F.connected8 F.connected9 F.connected10 regularization regularization regularization regularization regularization regularization regularization regularization regularization xT and yT are calculated using SlT and heading estimation
T as follows: xi = xi 1 + Sli 1 yi = yi 1 + Sli 1 cos( i 1) sin( i 1)
The same process is repeated for T2 and T3. During this process, the predicted locations converge to a small area. We calculate the centroid of these re ned locations which represent the user's predicted location Lp. 4
The experiments are conducted using Samsung Galaxy S8 (SM-G950N), and LG G6 (LGMG600L) devices. The features for training are extracted using Galaxy S8 collected data, while testing is done with two smartphones. The path used for the experiments is shown in Figure 5. The area of experiment building is 92 36 m2. Experiments are performed with people of di erent heights to evaluate the proposed approach.
Estimated locations at T1 Estimated locations at T2
Results are shown in Figure 6. Results demonstrate that the proposed approach can locate a person within 2.5 m, irrespective of the used device. Maximum errors are 13.33 m and 13.57 m for S8 and G6, respectively. However, the cumulative probability of maximum error is only 0.02. The mean error at 75% is 3.7 m and 4.1 m for S8 and G6. Even though the training is performed with S8 data, the localization errors are very similar for two smartphones. The underlying reason is the usage of features extracted from magnetic data than the magnetic data themselves. It also shows the usefulness of deep learning to assist in reliable indoor localization. We conducted additional experiments with less features also excluding 'interquartile', 'standard deviation', and 'mean absolute deviation'. Removing these features degrades the localization performance. 5
This research aims at using deep learning based ensemble classi er to solve indoor localization with heterogeneous devices. Three NNs are trained on magnetic data features from S8 smartphone. The experiments are performed with S8 and G6 smartphones. Results demonstrate that the use of features extracted from magnetic data are very fruitful to train NN. The localization accuracy is 2.5 m at 50% on two di erent devices. Mean error for S8 and G6 is 2.61 m and 2.95 m, respectively. Data collection is a laborious task which we intend to overcome with crowd sourcing data collection in future. How the larger indoor area may a ect the performance of the proposed approach is an intended future work.