A mean-offset classification technique was identified. It was found that the meanoffset classifier provides stability under dynamic indoor conditions and provides consistent results when training and test data combinations are swept from 10 - 95%. In this paper the meanoffset classifier is compared to the K-Nearest Neighbors (KNN) and Naïve Bayesian (NB) classifiers, with a view of developing an adaptable and computationally efficient indoor localization model using machine learning principles. It was seen that the mean-offset classifier improved results considerably and achieved an accuracy of 0.85 m and 1.15 m under line-of-sight (LOS) and non-line-of-sight (NLOS) conditions in residential areas.
1.1
There has been a rapid growth in localization techniques and the application thereof in
indoor environments [
There has been a growth in implementing machine learning algorithms to RSSI based
indoor localization. Table 1 summarizes the comparisons of localization using machine
learning techniques. In the Kernel-based learning method a spatial filtering step is
introduced to locate the estimated point to a subset of the environment and a kernelized
distance for estimating the Euclidean distance between the observed RSS and the stored
fingerprints is proposed, achieving an accuracy of 2.43m [
A mean-offset classifier which computes the percentage of error between the users RSSI against the database of trained RSSI centroids is implemented. The classifier is tested under dynamic indoor conditions where human movement is present and across multiple days to see if temporal fluctuations affect localization.
K-NN and Naïve Bayesian classifiers are implemented and compared to the mean-offset classifier under static indoor localization scenarios. Training data are swept from 10 – 90% to compare which classifier performs accurately when minimum training data are available.
Machine learning algorithms improve the accuracy of localization systems. All
machine learning algorithms are implemented in two phases, training and test phases [
norm of - dimensions RSSI vector , where
RSSI value from the database ̅ is represented as [
as the Euclidean norm-distance is used. The K-NN classifier
chooses the minimum distance of the K-neighbour points where is the list of points
given as
(̅ − ̅ ) ≤ (̅, ̅̅̅̅)
= { 1, … , }
̅̅1̅:̅̅ = {̅̅1̅, … , ̅̅̅}
̂ =
1
∑
=1
from the vector which must satisfy the following conditions [
The estimated location, ̂, by averaging the coordinated of the KNN classifier is 2.2
The Naïve Bayes classifier uses the Bayes theorem to makes classification decisions
with an assumption of conditional independence and uses conditional probabilities to
make classification decisions. In the indoor localization system, the location of a user
needs to be determined given the RSSI feature from APs. To achieve this the probability
of RSSI given in each region and the probability of the regions needs to be calculated.
The probability of the RSSI values is given by [
)|) ] ( | ( ∑1(| )∗ ( )∗ ( ))
)) ) are prior likelihood
distributions.
)are constant in all cases as there is no
in(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Where
( ∑1(|
[
( |
formation about the user’s position, hence making it a maximum likelihood estimate
= arg
[(|
)]
Hence, maximising the likelihood probability will help make a decision. This problem
is simplified to [
2( )| + 1 + |̅̅̅̅̅3̅̅̅̅−
3( )| 3 ( , ) = ( )
2 ) × 100] 2.3
grids, each grid with the physical location ( , )has a corresponding fingerprinting vector = { of centroids computed across each grid as follows: 1 1, 2 1, … , () 1)where ̅̅̅̅̅(̅̅)̅̅̅ is the database
̅̅̅̅̅̅̅1̅ = ∑ 1 =1 11 ( ) (11) (12) (13)
1( ), 2( ),
3 3.1
The indoor localization system was carried out on the second floor of a residential space. The dimension of the residential environment is 11 m by 5 m. It has three bedrooms (outlined in red, green and blue), a passage (outlined in purple) and one bathroom (outlined in orange), which are separated by walls and cupboards Fig. 2. There are 119 testing grids that are clustered as outlined in Fig. 2. In training, five RSSI measurements per AP, a total of 1785 RSSI measurements, were collected across all the grids. In the online phase, the user moved to five random locations and RSSI measurements from three APs were stored to a server via an API call and exported to Python, which applied the classifiers to locate the user’s position. Fig. 2. Floor map of the residential area. The blue, green, purple, orange and red outlines indicate the clustered areas. 3.2
To evaluate the classification techniques, the RSSI measurements collected during the training phase where applied to K-NN, NB and Centroid-Offset classifiers. The classifiers selected where provided training data that ranged from 10% - 90% as an efficient classifier needs to be able to adapt to scenarios where only minimum training data is available. The mean offset classifier outperforms the KNN and NB classifier when the split between the training-test data is between 20 % and 75% Fig. (3). As more training data are provided, the NB classifier marginally outperforms the mean-offset classifier. However, the NB classifier also struggles when < 60% training data are provided. Overall, the mean-offset classifier consistently achieves accuracy of within 1.3 m. The mean-offset classifier should account for dynamic changes presented by an indoor environment, such as temporal fluctuations in RSSI when recorded from the same point over multiple days and the movement of people and furniture. The Mean-Offset classifier was applied under two cases: tests were conducted on different times over multiple days and with movement of people.
Fig. 4. Measured RSSI values from five different locations over four days indicating the fluctuation experienced in RSSI values The test was conducted on five different locations over a period of four days. The fluctuations in RSSI across the three different APs was recorded and is represented in Fig. 4. The RSSI values had an average fluctuation of – 3dBm across the four days. The average accuracy over the locations at 12:00, 14:00, 18:00 and 20:00 was 1.08 m, 1 m, 0.83 m, 0.94 m and 0.58 m respectively Fig. 5. The accuracy over five locations across four days was 0.89 m. Mean-offset classifier was selected based on its performance as it shows adaptability under uncertain conditions and in instances where the environment is changing rapidly and the RSSI values are fluctuating. The test data gathered in an indoor residential area are applied to the mean-offset classifier and the results are tabulated in Table 2. Under LOS conditions an accuracy of 0.85 and 1.15 m was achieved in residential area.
In conclusion, the mean-offset classifier was proposed for fingerprint indoor localization using Wi-Fi. The mean-offset classifier computes the percentage of error between the users of RSSI against the database of trained RSSI centroids, after which the grid with the lowest percentage of error is chosen as the location implemented. Experimental results have demonstrated that the mean-offset classifier achieves an average localization of 0.86 m with over 70% of errors under 1 m, which outperforms other kernelbased, KNN, PCA-SVM, linear-SVM, gaussian-SVM, decision tree and random forest classifiers. Furthermore, main challenges such as RSSI instability and temporal ambiguity have been mitigated by the mean-offset classifier. The mean-offset classifier needs to be implemented for dynamic localization under several indoor locations such as malls, office spaces and underground parking. A hybrid between Artificial Neural Networks (ANN) and mean-offset is a concept that will be further investigated.