Page 184 - Kaleidoscope Academic Conference Proceedings 2024

2024 ITU Kaleidoscope Academic Conference

superior approach compared to others. learning algorithms have enhanced localization perfor-
In this approach, fingerprints, i.e. RSSI signal strength, mance, further improvements are still necessary. Exam-
are recorded in the Reference Points (RP) with the ples of these algorithms include DecisionTree [17], [18],
known locations, and this fingerprint database matches XGBoost [19], [20], and Random Forests [21], [22].
the pattern with the input signals to predict the position. The SRL-KNN algorithm [23] addresses the challenge
Furthermore, these approaches have high performance of limited mobility indoors by incorporating a penalty
compared to the geometric approach. The fingerprinting function that accounts for the physical distance between
approach has some drawbacks, such as a user’s current and previous positions when calculating
• Fingerprinting recording can be a tedious and time- fingerprint distances. This method effectively reduces
consuming process that requires significant effort. spatial ambiguity and improves performance.
• Localization accuracy improves if a large amount The H-LSMTF approach [9] improves localization
of fingerprints are recorded. accuracy by first removing outliers from the training data
• Most of the time, multiple access points are not to reduce output fluctuations caused by noise. It then
available. uses locally weighted regression (LWR) to further reduce
• Fingerprints need to be updated regularly in a uncertainty, followed by the integration of an encoder-
dynamic environment to ensure better accuracy. LSTM-encoder model. Evaluation on two datasets—one
• May not work effectively in an unobserved environ- in a 10m x 5m room (Dataset I) and another in a 35m x
ment and may require large fingerprints to achieve 16m laboratory (Dataset II)—showed improvements in
better results. localization accuracy of 20% and 60% for Dataset I and
Dataset II, respectively, compared to existing state-of-
Therefore, this paper uses dempster Shafer’s theory
the-art algorithms [9].
(DST) to predict regions using a single WiFi access
point. First, it converts the single RSSI values into The work [24] proposed an algorithm to address
degradation problems based on residual and long short-
multiple feature values, such as min, max, and quartiles
term (LSTM) networks. It is a spatial-temporal algorithm
of 25%, 50% and 75%. Then, it utilises the Dempster-
Shafer theory to predict the regions. Our main contribu- positioning that considers spatial features and temporal
sequential relations. The method uses a residual-based
tions are summarised as follows:
network to extract spatial characteristics of the Wi-Fi
1) The proposed algorithm works in a single WiFi
transmission and LSTM to record the temporal charac-
access point, requiring fewer fingerprints.
teristics of the Wi-Fi signal over a series of time slices.
2) The dataset was recorded using a smartphone in
A fully connected layer is then employed to estimate
a laboratory of 154 sq meters for a Single Access
the final location. This algorithm improves positioning
Point (AP). accuracy and reliability.
3) This paper uses the Dempster Shafer-Theory
Several techniques are utilized in indoor localiza-
(DST) approach to predict the regions.
tion to handle both imprecision and uncertainty. One
4) The proposed approach is compared with the ma-
prominent method is the Dempster-Shafer Theory (DST),
chine learning algorithms.
which was introduced by Dempster [25] in 1967 and
The remaining work is structured as follows: Section later formalized by Shafer [26] in 1976. This method es-
2 discusses the related works. Section 3 explains the timates the location of a target node by combining belief
dataset. Section 4 describes the proposed algorithm. Sec- masses from various positions, derived from probability
tion 5 summarizes the results and gives the conclusions. densities, using the PCR6 rule for evidence combination
in DST [27]. In this context, indoor environments are
II. RELATED WORKS
divided into zones separated by walls. The approach
RSSI-based indoor localization systems are cost- considers RSS (Received Signal Strength) irregularities
effective and often do not require extensive infrastructure by evaluating different distance intervals and the number
setup. These methods can be categorized into several of walls a radio signal traverses. These intervals are
types, including fingerprinting, geometric approaches, weighted based on an experimentally determined proba-
proximity methods, and others [15]. Among these, fin- bility density.
gerprinting has been particularly popular for many years. A recent study [28] proposes an innovative approach
In the fingerprinting approach, signal samples are ini- that integrates path loss algorithms with non-Bayesian
tially collected from specific locations, and during the data fusion based on DST. In this approach, belief
deployment phase, these signals are compared against masses are assigned to different positions within the
the predefined datasets. This method can use pattern- localization area based on RSS signals received from
matching or machine learning algorithms to estimate anchor nodes and the likelihood of the target node’s
locations [16]. Although machine learning and deep presence at these positions. The Log-Distance Path Loss

– 140 –