WiFi-Based Indoor Robot Positioning Using Deep

Fuzzy Forests
Le Zhang† , Zhenghua Chen† , Wei Cui∗ , Bing Li, Cen Chen, Zhiguang Cao, Kaizhou Gao

Abstract—Addressing the positioning problem of a mobile robot visual SLAM may suffer from various aspects, ranging from
remains challenging to date despite many years of research. the algorithmic point of view, such as the issues of camera cal-
Indoor robot positioning strategies developed in the literature ibration, data association, system initialization, and algorithms
either rely on sophisticated computer vision techniques to handle to compensate for camera distortions to hardware problems
visual inputs or require strong domain knowledge for non- including occlusions in the visual input. Moreover, the popular
visual sensors. Although some systems have been deployed, the
former may be lacking due to the intrinsic limitation of cameras
feature-based visual SLAM methods like [5] may fail when
(such as calibration, data association, system initialization, etc.) the scenes have blank background, e.g. indoor white walls.
and the latter usually only works under certain environment As an alternative, determining the robot’s location using non-
layouts and additional equipment. To cope with those issues, visual sensors such as radio frequency (RF) signals from radio
we design a lightweight indoor robot positioning system which sources is becoming popular. Examples include but are not
operates on cost-effective WiFi based Received Signal Strength limited to WiFi [6], [7], radio frequency identification device
(RSS) and could be readily pluggable into any existing WiFi (RFID) [8], [9], and ultra-wideband (UWB) [10]. Among those
network infrastructures. Moreover, a novel deep fuzzy forest approaches, a common practice is to use Received Signal
is proposed to inherit the merits of decision trees and deep Strength (RSS) measurements from radio sources in localizing
neural networks within an end-to-end trainable architecture. a mobile robot in the indoor/outdoor environment.
Real-world indoor localization experiments are conducted and
results demonstrate the superiority of the proposed method over In this paper, we study robots positioning using WiFi based
the existing approaches. signal strength. We choose WiFi due to its popularization in
residential buildings, academic areas, hospitals, and industry
Keywords—indoor robot positioning; WiFi; Deep fuzzy forests regions. More importantly, in comparison with other wireless
technologies, WiFi-based indoor localization is cost-effective
I. I NTRODUCTION because the system could be readily pluggable into any existing
Rapid developments in artificial intelligence have spurred WiFi network infrastructures including access points (APs),
the rise in robots service, in which robots positioning [1] cell phones, laptops, etc. We also focus on indoor scenarios
plays an inevitable role. For instance, using robots to perform because they are more complex. The line-of-sight (LOS)
programming tasks in harsh damaged environments, work as transmission channels between satellites and receivers is often
a pioneer in non-human living conditions such as the moon, invalid compared with their outdoor positioning counterparts
Mars, and other planets, take care of people in a nursing home in which mature localization systems such as global navigation
and deliver products in an indoor warehouse. However, robot satellite system (GNSS) has been developed successfully [11].
localization remains a challenging problem in practice despite Existing algorithms for RSS-based positioning can be typi-
many years of research. cally classified into two categories: 1) geometric related tech-
There are a plethora of research in the literature on using niques and 2) fingerprinting-based techniques [12]. Geometric
visual sensors to solve the simultaneous localization (“Posi- related techniques use theoretical and/or empirical path loss
tioning” is also named as “Localization” in the literature.) and models to characterize the RSS. Then, position estimation
mapping (SLAM) [2], [3], [4] problem of robots. However, of a node is performed using a trilateration or modeling
algorithm. One apparent drawback of these approaches, from
Copyright (c) 2020 IEEE. Personal use of this material is permitted. a practical point of view, is that in cluttered environments the
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to [email protected].
distance estimation is inaccurate because the attenuation of the
This work was supported by the National Natural Science Foundation of signal is poorly correlated with distance, especially under non-
China (61903231,61803104), Shandong Province Natural Science Foundation line-of-sight (NLOS) propagation [13]. The fingerprint-based
(ZR2018PF011), Singapore National Research Foundation (NRF-RSS2016- methods are becoming the dominating approach for indoor
004) and partially supported by the A*STAR Industrial Internet of Things positioning because wireless signal variances can be captured
Research Program under the RIE2020 IAF-PP Grant A1788a0023.
Le Zhang, Zhenghua Chen, Wei Cui and Cen Chen are with the Institute more accurately thanks to its transmissivity in complex indoor
for Infocomm Research, Agency for Science, Technology and Research spaces [14].
(A*STAR) Singapore. As illustrated in Fig. 1, a typical fingerprinting-based system
Bing Li is with the University of New South Wales, Australia. can be broken down into the offline phase and the online
Zhiguang Cao is with the National University of Singapore.
Kaizhou Gao is with the Macau University of Science and Technolog, P.R.C. phase. In the offline phase, an area of interest is calibrated
† indicates equal contribution and Wei Cui ([email protected]) is the by surveying the site with a mobile device (MD). The fin-
corresponding author. gerprints (i.e., the RSS values) from different predetermined
reference points (RPs) are collected as a priori knowledge Offline training phase
and recorded in a map database. During the online phase, RP1(x,y) RSS1,RSS2,…,RSSn
when an MD at unknown position sends its instant vector RP2(x,y) RSS1,RSS2,…,RSSn
of WiFi RSS values from surrounding APs, its current po- RP3(x,y) RSS1,RSS2,…,RSSn
sition will be estimated utilizing the information from the … … Training Database
map database. Compared with geometric related techniques, RPm(x,y) RSS1,RSS2,…,RSSn
fingerprinting-based methods have been widely regarded to be
more accurate. Hence, they have been heavily applied in real
life[15]. However, there are two major challenges of WiFi
fingerprinting-based technology for large-scale implementa-
RSS1,RSS2,…,RSSn Algorithm (x,y)
tion. One is to shrink the workload of the offline site survey
as the process is extremely laborious and time consuming.
The crowdsourcing based schemes have been proposed to Online localization phase
solve this problem. Instead of labeling the training database
manually, crowdsourcing could build the radio map with less Fig. 1: The offline phase and the online phase of a typical
maintenance effort by leveraging the unconscious participatory fingerprinting-based positioning.
sensing of mobile users [16], [17]. However, how to implement
crowdsourcing for the radio map is still an open problem
because the participant’s smartphones are usually different.
Another big issue is the vulnerability of RSS to environmental during the online phase are represented as a probability distri-
dynamics including variation in humidity, light, temperature, bution, and the algorithm estimates the position of the target
weather conditions[18], occupancy distribution, and human by calculating the highest probability of the mobile node’s
movement [19]. Hence the final performance of the system can location [27]. For instance, The authors in [28] proposed a
be dramatically degraded under a dynamic environment. This new algorithm based on Bayesian regression to estimate the
naturally opens a path towards designing more robust machine RSS posterior. Deterministic methods adopt machine learn-
learning techniques which can generalize well in practice. ing techniques to match the observed fingerprints with the
In this paper, we aim to solve the second aforementioned radio map database. More specifically, in [29], the authors
issue and develop a new light-weight indoor robot positioning proposed an indoor positioning system based on the smallest
system that uses low-cost WiFi devices which can be applied M-vertex polygon (SMP). Support Vector Machine/Regression
for various purposes, such as localization and navigation in (SVM/SVR) also has been widely applied for fingerprinting-
areas hard to reach by humans, navigational assistance for based position system. For example, Sun et al. [30] proposed
the blind, and tour guide robots. Motivated by the recent an indoor localization algorithm based on least-squares SVM
successes of random forest [20], [21], [22] and ensemble deep (LS-SVM). The work of [31] applied Support vector classi-
learning [23], [24] in different domains, a deep fuzzy forest is fication (SVC) for the indoor positioning problem. Wu et al.
designed which unifies decision trees with the representation [32] proposed to use the SVR method for position estima-
learning functionality known from deep neural networks in tion. Nerguizian [33] introduced the neural network model
an end-to-end trainable fashion. For simplicity, we adopt the and applied it for fingerprint based indoor positioning. A
commonly used deep fully-connected feed-forward network randomized network was introduced in [34]. As for deep
whose details are elaborated in Section V. The trees in our learning, a four-layer deep neural network was employed in
system work in a fuzzy manner to avoid wrong partitionings [35] for localization. However, it requires a relatively large
in internal nodes, which usually happens in conventional forest number of APs, which is less practical for indoor scenarios.
methods, and generalizes well in real-life scenarios. A deep semi-supervised reinforcement learning approach was
The rest of this paper is organized as follows. We discuss introduced [36] for indoor localization based on Bluetooth
the background and related works in Section II. Section III Low Energy (BLE) signal strength. Deep learning based indoor
describes the working process and localization scheme of fingerprinting systems using Channel State Information (CSI)
the proposed positioning system. Section IV explains the were proposed to achieve high localization precision [37], [38].
proposed deep fuzzy forest. The experimental evaluations are However, the transmitters in these system need to be equipped
demonstrated in Section V. Conclusion remarks are given in with CSI tools.
Section VI. The advantage of deterministic approaches is allowing for
automatically learning the potential relationship between the
RSS values and the locations, and most importantly, does not
II. R ELATED WORK
depend on knowing the entire environment layout upfront.
Nowdays, indoor localization based on smartphones has Due to similar reasons, machine learning based solutions saw
become a research focus due to its minimum hardware re- heavy usage across a wide range of fields and tasks, examples
quirement [25], [26]. Generally speaking, fingerprinting based including: computer vision [24], speech recognition [39], re-
on RSS techniques can be generally categorized into two inforcement learning [40] and affective computing [41]. The
groups: probabilistic approaches and deterministic approaches. unifying idea behind all of the above is deep learning, the
For probabilistic approaches, observed signal strength vectors utilisation of neural networks with many hidden layers, for
the purposes of learning complex feature representations from
raw data, rather than relying on handcrafted feature extraction.
In addition, recent works on ensemble learning [23], [22] also
show that using multiple models could effectively improve
the generalization ability of the learning system. Motivated
by those work, a deep fuzzy forest is proposed in this paper
to inherit the merits of ensemble methods and deep neural
networks within an end-to-end trainable architecture. Extensive
real-world indoor localization experiments are conducted and
results demonstrate the superiority of the proposed method
over the existing approaches. Fig. 3: Block diagram of hardware mounted on the robot.
It consists of a development board with an ARM processor,
III. S YSTEM D ESIGN microcontrollers, a wireless network card and a motor driver.
A. System Architecture
The whole system architecture of the proposed positioning
system for the robot is illustrated in Fig. 2. The mobile robot B. Hardware Design of the robot
is equipped with a commercial radio transceiver to extract RSS
The hardware mounted on the robot with its detailed com-
data from all APs. The RSS signals from each AP are encoded
ponents is presented in Fig. 3. It consists of a development
as a data packet and further forwarded to the back-end server
board with a 1GHz Samsung S5P V 210 ARM Cortex − A8
for both radio map construction and learning. In this paper,
processor, an ST M 32F 103 chip (a version of ST M 32 micro-
we only focus on the localization part and its key components
controllers with an ARM Cortex − M 3 core and a maximum
are elaborated in the following section.
frequency of 72M Hz), an RT 3070 wireless network card and
the L298N motor driver. The RT 3070 card is used to com-
municate with the server for sending/receiving collected RSS
packets. The ST M 32F 103 chip is employed to interact with
users for motion control through the universal asynchronous
receiver/transmitter (UART) protocol. The weakest detectable
RSS is set to be −95dBm.

IV. P ROPOSED P OSITIONING A PPROACH

A. WiFi Indoor Positioning Problem Setting
Given m RPs and n APs with labels, the system architecture
of the proposed indoor positioning system (IPS) consists of
two stages: offline radio map database construction and online
localization. In the first stage, RSS samples from m RPs are
collected by the mobile robot from n APs for the construction
of the fingerprinting database. The received RSS values at RPs
Fig. 2: The system architecture of our system. It consists from all APs are denoted S = {S1 , . . . , Si , . . . , Sm } ∈ IRn×m ,
of four main components: a mobile robot, commercial off- where Si represents the RSS values at i-th RP from all APs.1
the-shelf WiFi access points (APs), a server and a web- Each fingerprint sample is a vector of RSS of n dimension and
based monitoring system. In the first stage, RSS samples
is denoted as Si = [s1i , ..., sji , ..., sni ]> . More formally, Si ∈
S = {S1 , . . . , Si , . . . , Sm } ∈ IRn×m from m Reference Points
S ⊆ IRn , i ∈ {1, . . . , m}. The corresponding fingerprinting
(demoted by the red points) are collected by the mobile robot
dataset consists of (Pi , Si )m i=1 , Pi ∈ P, where Pi = (xi , yi ) is
from n Access Points (illustrated by the routers here) for
the coordinate of the i-th RP.
the construction of the fingerprinting database. The database
In the second stage, a set of signal strength measured
consists of (Pi , Si )m i=1 , Pi ∈ P, where Pi = (xi , yi ) is from n APs at the testing point (TP) is denoted St =
the coordinate of the i-th RP. We learn a mapping function
G : S → P, which is a deep fuzzy random forest as elaborated [s1t , . . . , sjt , . . . , snt ]> , where sjt is the mean signals strength
in Section IV-B, on the server in an off-line manner to regress value received from APj during a period of time. We solve
the location of the robot based on the RSS inputs. After the localization problem in this setting by estimating the
that, real-time location inference is conducted by using the corresponding coordinate Pt = (xt , yt ) of fingerprinting vector
mapping function G with the new received RSS samples and St with offline database. It could be achieved by learning a
the web-based monitoring system could then make necessary mapping function G : S → P. The key problem of indoor
adjustment to control the trajectory of the robot. 1 For the simplicity of notation, here we assume that each RP is associated
with one RSS sample. However, in practice, we may generate multiple samples
over a period of time.
localization in our case is the reformulated as learning the the overall loss of the k th deep fuzzy tree gk at internal node
mapping function G with the offline fingerprinting database. j is:
n
B. Deep Fuzzy Forest XX
L(gk ) = Pkj (Si )(||f ckj (N (Si ) − Pi ||2
Methodology: We consider our mapping function G as a deep j∈L i=1
fuzzy random forest, composed of K base deep fuzzy tree + ||f cj (N (Si )) − Pi ||) (2)
regressors G = {gk }K k=1 , where the regressor gk : S → X
P, k ∈ {1, . . . , K}, called fuzzy decision tree are combined +λ∗ Ldis
j (gk ),
using averaging. Each decision tree gk (N (S)) regresses the j∈I
label of a sample Si ∈ S (i = 1, 2, ·, m) by firstly extracting
high-level abstract features from a network N (please refer where pkj (Si ) is the probability of assigning the Si to the
to the experiment section for the network structure) and then leaf node j in the k th tree, which is accumulated along its
routing it from the root to some leaf node, recursively, which own routing path, as illustrated in Fig. 4. The final location of
provides a label for the instance. Specifically, each node j each sample is obtained by a weighted average of the results
(j ∈ I where I stands for all the index of the internal with its corresponding probability from each leaf node. All
nodes) in the tree is associated with a fuzzy split function the parameters, including the feature extracting part N , the
fkj (N (Si ), θkj ) (where θkj is the parameter of the split splitting function fj (jL) as well as the regression function
function.) defined on fuzzy sets Al and Ar which stand for in the leaf node f cj (j ∈ L) are jointly optimized during the
assigning the data to the left and right child node, respectively. back-propagation.
In our setting each internal node of the tree, whose division is
based on the results of fkj (N (Si ), θkj ), which is embodied Differences with existing methods. Indeed, we are certainly
by a fully-connected layer therein, generates a child node not the first to propose deep random forests. For instance,
for each fuzzy set of the partition. The fuzzy partition of in [43], a deep neural forest was proposed for image clas-
each internal node guarantees completeness (no point in the sification. Moreover, Shen et al. introduced a deep regres-
domain is outside of the fuzzy partition), and is a strong fuzzy sion forest [44] for age estimation. In addition, Zhou et al.
partition, i.e., fkj (N (Si ), θkj ) ∈ {pkj (Si ), 1−pkj (Si )}, where introduced the gcForest [45] method which can be trained
pkj (Si ) ∈ [0, 1] stands for the probability of assigning the data without the back-propagation. While we take inspiration from
sample Si to the leaf child node. Given an input Si , the output them, we are the first to apply deep forest for indoor robot
of the tree is the prediction stored at the leaf reached by Si , localization. Moreover, all the parameters in our methods are
which is a target label P ∈ P in our case. jointly optimized by back-propagation while the internal and
In a decision tree, each split function at a non-leaf node leaf nodes in both the work [27], [44] are optimized in an
separates the data so that the further separations in the children iterative manner. And the forest in [45] are optimized in a
nodes are easier. Under this umbrella, conventional decision layer-wise fashion without back-propagation.
trees such as CART [42] utilize heuristic impurity scores to
guide the splitting process of the data samples such that the
data samples reaching the child node is purer. Motivated by V. E XPERIMENTAL E VALUATIONS
this, we further propose a new distribution loss to purse a low
purity in each child node after splitting. More specifically, our In order to evaluate the performance of the proposed posi-
distribution loss for the k th deep fuzzy tree gk is formulated tioning system, we conduct experiments in an indoor labora-
as: tory environment with an area of 512m2 (16m × 32m) using
Xn the aforementioned WiFi positioning robot. The dimensional
Ldis
j (g k ) = pkj (Si ) ∗ ||N (Si ) − cj ||2 , and drawing of the environment is shown in Fig. 5. In our setting, 8
i=1 TP-LINK W DR6500 Routers fixed on 1.5-meter-high tripods
n (1) were adopted as APs in our experiments.
X
cj = pkj (Si ) ∗ N (Si ) We select 133 different points to collect RSS fingerprints of
i=1 the mobile robot, including 107 offline calibration points and
The loss in Eqn (1) naturally reduces the intra-cluster 26 online testing points, as presented in Fig. 5. At each point,
variations. It helps to learn a better split function fkj to RSS information from 500 packet receptions is collected. Thus,
generate a purer partitioning for each child node which eases 107 × 500 × 8 RSS information was used to train the proposed
the future optimization therein. In addition, in many scenarios method. The grid size between two adjacent training points
where the noisy input data may not be correctly assigned to is set to be roughly 1.5m. Due to the environment dynamics
one of the child node, our fuzzy partitioning function works have a large impact on the RSS measurements, the offline data
in a fuzzy way such that this error can be further amended in collection is performed in different periods of a month.
the children nodes. For the performance criteria, we adopt the widely used mean
For the leaf nodes j ∈ L (where L represents all the index square error (MSE), mean absolute error (MAE) and Root
of the leaf nodes), we also use a standard fully-connected layer Mean Squared Error (RMSE). Denote the predicted value of
f c to regress the labels with each input data samples. Hence, a model and the ground truth target by P̃ and P , respectively,
 Decision Forest

Tree

DNN

……
Tree
S1(s1,s2 ,...,sn)
P1(x1,y1)
S2(s1,s2 ,...,sn), P2(x2,y2)
.
.
.
.
.
.
Sm(s1,s2 ,...,sn) Pm(x1,ym)

……
Tree

Fig. 4: System overview of the proposed deep fuzzy forest. Each tree received highly abstract features from a deep network
and divided the data samples recursively until certain criteria are met. More specifically, considering a single tree, each decision
node is implemented with a fully-connected layer with a Sigmoid activation function and its output denotes the corresponding
probability of assigning the current datum to the left and right child node, respectively. Please refer to Section V for more details
on the network structure. The final results are obtained by weighted averaging.

mean square error can be defined as:

5m 1m 6m 1m 19m
M SE = avg((P̃ − P )2 )
M AE = avg(|P̃ − P |) (3) 4m 4m

2
RM SE = sqrt(avg((P̃ − P ) )
4m
where avg(·) is the average over all the test samples.
We coin our method as DeepFuzzy and compare it with 10m
other widely used regression methods, including (1) k nearest
neighbours [46], (2) Random Forest [47], (3) SVR [48], 8m

(4) Randomized Neural Networks [34], (5) Deep Regression

2m
Forest [44], (6) Deep Network, which uses the same network
11m 3m 10m 2m 6m
backbone N as done in DeepFuzzy and a conventional fully-
connected layer for regression, and (6) DeepFuzzy-NoDis, TP-LINK WDR6500 Router Training Positions Testing Positions

which ignores the distribution loss by setting the λ in Eqn (2)

Fig. 5: The layout of the testbed. An indoor laboratory with
to be zero; We do not compare it with either gcForest [45] or
an area of 512m2 (16m × 32m) and 8 TP-LINK W DR6500
Deep Neural Forest [43] because they are originally designed
Routers fixed on 1.5-meter-high tripods were adopted as APs
for classification problems. For all methods, we optimize their
in our experiments. There are 133 different points and the grid
hyper-parameters on the training set. We implemented our
size is set to be roughly 1.5m.
DeepFuzzy in Pytorch and trained using the SGD optimizer
with an initial learning rate of 1e − 5. The batch size is set
to be 200 and trained with a maximum epoch of 85. The
maximum tree depth is set to be 6. In each node we use a TABLE I: Performance comparison of different approaches
50% randomly selected subset of the input. λ in Eqn( 2) We for our indoor robot localization. DeepFuzzy achieves the best
found our method could be able to reach a saturating-well performance with all metrics.
performance. See Section V-B for more details. Our system
is light-weighted and could be used for real-time application. Method MSE MAE RMSE
More specifically, it takes 4.298 seconds for a whole epoch. KNN [46] 7.60 1.86 2.76
Random Forest [47] 6.23 1.64 2.50
For testing, it takes 0.03 ms to estimate the location for a SVR [48] 4.86 1.62 2.20
single datum. All the experiments are conducted in Ubuntu Randomized Network [34] 14.20 2.89 3.77
Deep Regression Forest [44] 6.23 1.64 2.49
14.04 with a 12 GB Titan-X GPU and an Intel i7 CPU (3.4 Deep Network 4.63 1.73 2.15
GHZ). All the experiments are conducted in Ubuntu 14.04 with DeepFuzzy-NoDis 4.58 1.63 2.14
a 12 GB Titan-X GPU and an Intel i7 CPU (3.4 GHZ). For DeepFuzzy 3.20 1.36 1.79
the learning curves, please refer to Fig. 7.
As for the feature extraction network N , we design it in
such a way that with a conventional fully connected layer, may make the optimization of DeepFuzzy more challenging.
it achieves the best performance on the training set. More As for the λ in Eqn (2), we empirically found that our system
specifically, we use a standard 4-layer feed-forward network. could yield comparably good results when setting it to be
The network receives 8 RSS values as input and generates within [0.01, 0.05]. Due to page limits, we omit the graphic
128,256,512 and 1024 outputs in the following layers, respec- illustration.
tively. We use ReLU in each layer as the activation function 2) The grid size: As explained in Section V and Fig. 5, the
and use a dropout layer with a dropout probability of 0.3 before grid size actually represents the precision of the system by
feeding the features into the trees to reduce the risk of over- directly controlling the number of training samples. Setting a
fitting. smaller grid size may involve more training points in Fig. 5
for the model to learn. To verify this, we design a simple
experiment in which we manually control the grid size of our
A. Performance Comparison
environment by removing part of the training samples from
Table I shows the comparison of different approaches on the training set. More specifically, we set the grid size to be
our dataset. We observe the following phenomena (1) Deep {1.5, 3, 4.5} and the results are shown in Table II. It can be
learning approaches, such as Deep Regression Forest [44], seen that the error (MAE) of our system scales up linearly with
Deep Network and DeepFuzzy, generally outperform other the actual value of grid size. Hence, collecting a large scale
non-deep methods such as Random Forests and SVR. (2)
DeepFuzzy achieves the best performance regarding all the
metrics, which suggests the effectiveness of the proposed TABLE II: The effectiveness of the grid size. The MAE of
method. (3) DeepFuzzy variants (DeepFuzzy-NoDis and Deep- DeepFuzzy scales up linearly with the grid size.
Fuzzy) outperform the same network backbone. This clearly Grid Size MSE MAE
indicates the feasibility and superiority of the proposed fuzzy 1.5 3.20 1.36
forest compared with the conventional fully connected layer. 3 10.02 2.97
4.5 27.32 4.63
One may argue that the better performance of DeepFuzzy may
be caused by a larger model capacity as more parameters are
dataset, as done in other AI problems, seems to be an effective
involved. To answer this, we also design an experiment to
way to improve our final localization performance. However,
firstly project the features from the feature extraction network
on one hand, manual labeling is costly, time-consuming, error-
backbone N into a higher dimension with the same amount of
prone and requires massive human intervention. This is of-
parameters of the best configuration of DeepFuzzy. Interesting,
ten impractical, especially for our indoor robot localization
in this setting, a slightly worse performance is observed. (4)
problem because the temporal sampling rate of RSS data is
DeepFuzzy generally performs better than DeepFuzzy-NoDis,
low (e.g.,2-5 RSS samples per second). On the other hand,
which directly shows that regularizing the network towards
suffering from the dynamics in the environment, if any two
purer partitionings, as defined in Eqn (1), improves the final
training points are closer together, the distributions of their
performance. The predicted trajectories of training points and
RSS features are highly overlapping in practice. Based on the
testing points with trained DeepFuzzy are illustrated in Fig.9.
analysis in [49], the recommended grid spacing between two
adjacent locations should be larger than 1.25m.
B. Ablation Study 3) The ensemble size: Generally speaking, the performance
Apart from the distribution loss which has been included of a typical ensemble system will not become worse if we
in Table I, in this section we further investigate the following increase K [47]. However, our DeepFuzzy is slightly different
aspects of DeepFuzzy. from conventional random forest [47] where each tree was
1) The tree depth and λ: To show this, we set the maximum trained independently. Instead, in order to further reduce the
tree depth to be d (d ∈ {0, 2, 4, 6, 8, 12}). Results are illus- risk of over-fitting as well as mode complexity, the feature
trated in Figure 6. We empirically observe a “U-like” shape of extraction network backbone N is shared amongst all the
a trend in which a smaller value of tree depth is not sufficient ensemble members. During our preliminary study, we empiri-
to utilize the fuzzy merits of DeepFuzzy while a larger value cally find that this strategy helps to train DeepFuzzy faster and
 4m 4m

4m

10m

8m

2m
11m 3m 10m 2m 6m

TP-LINK WDR6500 Router Training Positions Predicted Training Positions

Fig. 6: The performance of DeepFuzzy with respective to the
5m 1m 6m 1m 19m
value of tree depth. We empirically observe a “U-like” shape
of trend in which a smaller value of tree depth is not sufficient
4m 4m
to utilize the fuzzy merits of DeepFuzzy while a larger value
may make the optimization of DeepFuzzy more challenging.
4m

Learning Curves
50 10m
Training Loss
Testing Loss 8m
40

30 2m
MSE

11m 3m 10m 2m 6m
20
TP-LINK WDR6500 Router Testing Positions Predicted Testing Positions

10
Fig. 9: The predicted trajectories of training points and testing
0 points with trained DeepFuzzy.
0 20 40 60 80 100
Epoch

Fig. 7: The learning curves of the proposed method.

extra experimental results with different number of APs and
Packages and results are summarized in Table III. Results
clearly show that using more APs and Packages lead to better
better, and more interesting, this sharing strategy also helps to performances.
regularize the network and achieves satisfactory results with a
smaller ensemble size. See Fig. 8 for more details.
4) The number of APs and Packets: Our experimental TABLE III: Effect of the number of APs and Packets.
environments adopt 8 APs and 500 Packets. In order to
investigate the effectiveness of those factors, we have done (a) Effect of the number of APs. (b) Effect of the number of Packets.

Number of APs MSE MAE Number of Packets MSE MAE

8 3.20 1.36 500 3.20 1.36
4 7.43 1.99 250 3.55 1.43

5) The depth of network: In our proposed method, we

employed a 4-layer network. In order to explore the impact
of the depth of network, comprehensive experiments have
been conducted with different network structures, results are
reported in Table IV. The 4-layer network (“4 Layers +
Forest”) receives 8 RSS values as input and generates 128,
Fig. 8: The performance of DeepFuzzy with respect to the 256, 512 and 1024 outputs in the following layers, respectively.
ensemble size K. We empirically find that DeepFuzzy may The network of “3 Layers + Forest” stands for the case where
achieve satisfactory results with a relatively smaller ensemble the network receives 8 RSS values as input and generates 256,
size. 512 and 1024 outputs in the following layers, respectively.
Moreover, the network of “5 Layers + Forest” stands for the
case where the network receives 8 RSS values as input and [12] C. Wu, Z. Yang, Y. Liu, and W. Xi, “Will: Wireless indoor localization
generates 128, 256, 256, 512 and 1024 outputs in the following without site survey,” IEEE Transactions on Parallel and Distributed
Systems, vol. 24, no. 4, pp. 839–848, 2013.
layers, respectively. The results demonstrate that the 4-layer
network outperform others. [13] D. Dardari, P. Closas, and P. M. Djurić, “Indoor tracking: Theory, meth-
ods, and technologies,” IEEE Transactions on Vehicular Technology,
vol. 64, no. 4, pp. 1263–1278, 2015.
TABLE IV: Effect of the depth of network. [14] F. Zhao, H. Luo, X. Zhao, Z. Pang, and H. Park, “Hyfi: Hybrid floor
identification based on wireless fingerprinting and barometric pressure,”
Network Structure MSE MAE IEEE Transactions on Industrial Informatics, vol. 13, no. 1, pp. 330–
3 Layers + Forest 4.47 1.63 341, 2017.
4 Layers + Forest 3.20 1.36
5 Layers + Forest 3.98 1.43 [15] Z. Yang, C. Wu, and Y. Liu, “Locating in fingerprint space: wireless
indoor localization with little human intervention,” in Proceedings of
the 18th annual international conference on Mobile computing and
networking. ACM, 2012, pp. 269–280.
VI. C ONCLUSION [16] A. Rai, K. K. Chintalapudi, V. N. Padmanabhan, and R. Sen, “Zee: Zero-
A lightweight indoor robot localization system is developed effort crowdsourcing for indoor localization,” in Proceedings of the 18th
annual international conference on Mobile computing and networking.
in this paper using WiFi-RSS as input. Our system is generic ACM, 2012, pp. 293–304.
and could be readily pluggable into any existing WiFi network [17] C. Wu, Z. Yang, and Y. Liu, “Smartphones based crowdsourcing for
infrastructures. A novel end-to-end deep fuzzy forest is intro- indoor localization,” IEEE Transactions on Mobile Computing, vol. 14,
duced for robust location estimation. We conduct experiments no. 2, pp. 444–457, 2015.
in real-world scenarios and results establish the superiority [18] J. Luomala and I. Hakala, “Effects of temperature and humidity on radio
of the proposed deep random forest method over the existing signal strength in outdoor wireless sensor networks,” in 2015 Federated
machine learning approaches. Conference on Computer Science and Information Systems (FedCSIS).
IEEE, 2015, pp. 1247–1255.
[19] C. Yang and H.-R. Shao, “Wifi-based indoor positioning,” IEEE Com-
R EFERENCES munications Magazine, vol. 53, no. 3, pp. 150–157, 2015.
[1] M. S. Miah, J. Knoll, and K. Hevrdejs, “Intelligent range-only mapping [20] L. Zhang and P. N. Suganthan, “Oblique decision tree ensemble via
and navigation for mobile robots,” IEEE Transactions on Industrial multisurface proximal support vector machine,” IEEE transactions on
Informatics, vol. 14, no. 3, pp. 1164–1174, 2018. cybernetics, vol. 45, no. 10, pp. 2165–2176, 2014.
[2] J. Zhang, M. Gui, Q. Wang, R. Liu, J. Xu, and S. Chen, “Hierarchical [21] L. Zhang, J. Varadarajan, P. Nagaratnam Suganthan, N. Ahuja, and
topic model based object association for semantic slam,” IEEE trans- P. Moulin, “Robust visual tracking using oblique random forests,” in
actions on visualization and computer graphics, vol. 25, no. 11, pp. Proceedings of the IEEE Conference on Computer Vision and Pattern
3052–3062, 2019. Recognition, 2017, pp. 5589–5598.
[3] G. Bresson, Z. Alsayed, L. Yu, and S. Glaser, “Simultaneous localiza-
[22] L. Zhang and P. N. Suganthan, “Benchmarking ensemble classifiers with
tion and mapping: A survey of current trends in autonomous driving,”
novel co-trained kernel ridge regression and random vector functional
IEEE Transactions on Intelligent Vehicles, vol. 20, pp. 1–1, 2017.
link ensembles [research frontier],” IEEE Computational Intelligence
[4] A. J. Davison, I. D. Reid, N. D. Molton, and O. Stasse, “Monoslam: Magazine, vol. 12, no. 4, pp. 61–72, 2017.
Real-time single camera slam,” IEEE Transactions on Pattern Analysis
& Machine Intelligence, no. 6, pp. 1052–1067, 2007. [23] L. Zhang, Z. Shi, M.-M. Cheng, Y. Liu, J.-W. Bian, J. T. Zhou,
G. Zheng, and Z. Zeng, “Nonlinear regression via deep negative corre-
[5] R. Mur-Artal, J. M. M. Montiel, and J. D. Tardos, “ORB-SLAM: a lation learning,” IEEE Transactions on Pattern Analysis and Machine
versatile and accurate monocular SLAM system,” IEEE transactions Intelligence, 2019.
on robotics, vol. 31, no. 5, pp. 1147–1163, 2015.
[24] L. Zhang, Z. Shi, J. T. Zhou, M.-M. Cheng, Y. Liu, J.-W. Bian, Z. Zeng,
[6] C. Kumar and K. Rajawat, “Dictionary-based statistical fingerprinting and C. Shen, “Ordered or orderless: A revisit for video based person
for indoor localization,” IEEE Transactions on Vehicular Technology, re-identification,” IEEE Transactions on Pattern Analysis and Machine
vol. 68, no. 9, pp. 8827–8841, 2019. Intelligence, 2020.
[7] C. Feng, W. S. A. Au, S. Valaee, and Z. Tan, “Received-signal-
[25] X. Tian, Z. Song, B. Jiang, Y. Zhang, T. Yu, and X. Wang, “Hiquadloc:
strength-based indoor positioning using compressive sensing,” IEEE
A rss fingerprinting based indoor localization system for quadrotors,”
Transactions on Mobile Computing, vol. 11, no. 12, pp. 1983–1993,
IEEE Transactions on Mobile Computing, vol. 16, no. 9, pp. 2545–2559,
2012.
2016.
[8] D. Hahnel, W. Burgard, D. Fox, K. Fishkin, and M. Philipose, “Mapping
and localization with rfid technology,” in Robotics and Automation, [26] H. Liu, J. Yang, S. Sidhom, Y. Wang, Y. Chen, and F. Ye, “Accurate
2004. Proceedings. ICRA’04. 2004 IEEE International Conference on, wifi based localization for smartphones using peer assistance,” IEEE
vol. 1. IEEE, 2004, pp. 1015–1020. Transactions on Mobile Computing, vol. 13, no. 10, pp. 2199–2214,
2013.
[9] E. DiGiampaolo and F. Martinelli, “Mobile robot localization using
the phase of passive uhf rfid signals,” IEEE Transactions on Industrial [27] P. Kontkanen, P. Myllymäki, T. Roos, H. Tirri, K. Valtonen, and
Electronics, vol. 61, no. 1, pp. 365–376, 2014. H. Wettig, “Probabilistic methods for location estimation in wireless
networks,” in Emerging Location Aware Broadband Wireless Ad Hoc
[10] M. Z. Win and R. A. Scholtz, “Characterization of ultra-wide bandwidth
Networks. Springer, 2005, pp. 173–188.
wireless indoor channels: A communication-theoretic view,” IEEE Jour-
nal on Selected Areas in Communications, vol. 20, no. 9, pp. 1613– [28] M. M. Atia, A. Noureldin, and M. J. Korenberg, “Dynamic online-
1627, 2002. calibrated radio maps for indoor positioning in wireless local area
[11] S.-C. Yeh, W.-H. Hsu, M.-Y. Su, C.-H. Chen, and K.-H. Liu, “A networks,” IEEE Transactions on Mobile Computing, vol. 12, no. 9,
study on outdoor positioning technology using gps and wifi networks,” pp. 1774–1787, 2013.
in Networking, Sensing and Control, 2009. ICNSC’09. International [29] P. Prasithsangaree, P. Krishnamurthy, and P. Chrysanthis, “On indoor
Conference on. IEEE, 2009, pp. 597–601. position location with wireless lans,” in Personal, Indoor and Mobile
 Radio Communications, 2002. The 13th IEEE International Symposium
on, vol. 2. IEEE, 2002, pp. 720–724.
[30] G. Sun and W. Guo, “Robust mobile geo-location algorithm based on
ls-svm,” IEEE Transactions on vehicular technology, vol. 54, no. 3, pp.
1037–1041, 2005.
[31] C.-L. Wu, L.-C. Fu, and F.-L. Lian, “WLAN location determination in
e-home via support vector classification,” in Networking, sensing and
control, 2004 IEEE international conference on, vol. 2. IEEE, 2004,
pp. 1026–1031.
[32] Z.-l. Wu, C.-h. Li, J. K.-Y. Ng, and K. R. Leung, “Location estimation
via support vector regression,” IEEE Transactions on mobile computing,
vol. 6, no. 3, 2007.
[33] C. Nerguizian, C. Despins, and S. Affès, “Geolocation in mines with an
impulse response fingerprinting technique and neural networks,” IEEE
Transactions on Wireless Communications, vol. 5, no. 3, pp. 603–611,
2006.
[34] W. Cui, L. Zhang, B. Li, J. Guo, W. Meng, H. Wang, and L. Xie, “Re-
ceived signal strength based indoor positioning using a random vector
functional link network,” IEEE Transactions on Industrial Informatics,
vol. 14, no. 5, pp. 1846–1855, 2018.
[35] W. Zhang, K. Liu, W. Zhang, Y. Zhang, and J. Gu, “Deep neural
networks for wireless localization in indoor and outdoor environments,”
Neurocomputing, vol. 194, pp. 279–287, 2016.
[36] M. Mohammadi, A. Al-Fuqaha, M. Guizani, and J.-S. Oh, “Semisu-
pervised deep reinforcement learning in support of iot and smart city
services,” IEEE Internet of Things Journal, vol. 5, no. 2, pp. 624–635,
2018.
[37] X. Wang, L. Gao, and S. Mao, “CSI phase fingerprinting for indoor
localization with a deep learning approach,” IEEE Internet of Things
Journal, vol. 3, no. 6, pp. 1113–1123, 2016.
[38] X. Wang, L. Gao, S. Mao, and S. Pandey, “CSI-based fingerprinting for
indoor localization: A deep learning approach,” IEEE Transactions on
Vehicular Technology, vol. 66, no. 1, pp. 763–776, 2016.
[39] D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg,
C. Case, J. Casper, B. Catanzaro, Q. Cheng, G. Chen et al., “Deep
speech 2: End-to-end speech recognition in english and mandarin,” in
International Conference on Machine Learning, 2016, pp. 173–182.
[40] Z. Cao, H. Guo, W. Song, K. Gao, Z. Chen, L. Zhang, and X. Zhang,
“Using reinforcement learning to minimize the probability of delay oc-
currence in transportation,” IEEE Transactions on Vehicular Technology,
vol. 69, no. 3.
[41] L. Zhang, S. Peng, and S. Winkler, “PersEmoN: A deep network for
joint analysis of apparent personality, emotion and their relationship,”
IEEE Transactions on Affective Computing, 2019.
[42] L. Breiman, Classification and regression trees. Routledge, 2017.
[43] P. Kontschieder, M. Fiterau, A. Criminisi, and S. Rota Bulo, “Deep
neural decision forests,” in Proceedings of the IEEE international
conference on computer vision, 2015, pp. 1467–1475.
[44] W. Shen, Y. Guo, Y. Wang, K. Zhao, B. Wang, and A. L. Yuille,
“Deep regression forests for age estimation,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2018,
pp. 2304–2313.
[45] Z.-H. Zhou and J. Feng, “Deep forest: Towards an alternative to deep
neural networks,” arXiv preprint arXiv:1702.08835, 2017.
[46] N. S. Altman, “An introduction to kernel and nearest-neighbor non-
parametric regression,” The American Statistician, vol. 46, no. 3, pp.
175–185, 1992.
[47] L. Breiman, “Random forests,” Machine learning, vol. 45, no. 1, pp.
5–32, 2001.
[48] A. J. Smola and B. Schölkopf, “A tutorial on support vector regression,”
Statistics and computing, vol. 14, no. 3, pp. 199–222, 2004.
[49] K. Kaemarungsi and P. Krishnamurthy, “Analysis of WLAN’s received
signal strength indication for indoor location fingerprinting,” Pervasive
and mobile computing, vol. 8, no. 2, pp. 292–316, 2012.