Vol-3194/paper48

Paper

Paper
edit
description
id	Vol-3194/paper48
wikidataid	Q117344933→Q117344933
title	Supporting Big Moving Objects Tracking and Analysis: An Innovative Scan-Matching Algorithm
pdfUrl	https://ceur-ws.org/Vol-3194/paper48.pdf
dblpUrl	https://dblp.org/rec/conf/sebd/CuzzocreaSLM22
volume	Vol-3194→Vol-3194
session	→

Supporting Big Moving Objects Tracking and Analysis: An Innovative Scan-Matching Algorithm

Supporting Big Moving Objects Tracking and
Analysis: An Innovative Scan-Matching Algorithm⋆
Alfredo Cuzzocrea1,2,* , Majid Abbasi Sisara1,3 , Kristijan Lenac4 and Enzo Mumolo3
1
  iDEA Lab, University of Calabria, Rende, Italy
2
  LORIA, University of Lorraine, Nancy, France
3
  DIA Dept., University of Trieste, Trieste, Italy
4
  Faculty of Engineering, University of Rijeka & Center for Artificial Intelligence and Cybersecurity, Rijeka, Croatia


                                         Abstract
                                         In this paper we propose a novel family of scan matching algorithms based on registration, which are
                                         enhanced by using a genetic pre-alignment phase based on a novel metrics, fist, and, second, performing
                                         a finer alignment using a deterministic approach. Our experimental assessment and analysis confirms
                                         the benefits deriving from the proposed novel family of such algorithms.

                                         Keywords
                                         Moving Objects, Scan-Matching Algorithms, Intelligent Systems, Genetic Optimization




1. Introduction
Big moving objects arise as a novel class of big data objects in emerging environments. Here,
the main problems are the following: (i) tracking, which represents the baseline operation for
a plethora of higher-level functionalities, such as detection, classification, and so forth; (ii)
analysis, which meaningfully marries with big data analytics scenarios.
   Nowadays, a great deal of interest is growing around the mobile object tracking problem,
especially due to the emerging integration between robotics and big data applications (e.g.,
[8, 25, 12]). Following this trend, several mobile object tracking approaches have recently
appeared in literature, considering different aspects of the target issue, such as coverage,
completeness, effectiveness, efficiency, etc. The category of algorithms that goes under the name
of scan-matching (e.g., [22, 20, 16]) supports mobile objects positioning in indoor environments
based on the acquisition of maps of the environment surrounding the target mobile objects.
Maps are acquired from two successive points in the objects’ path using a range-scanner sensor
positioned on mobile objects themselves. The first acquisition is called reference scan and the
second actual scan. The actual scan is sometimes also called new scan. By overlapping the maps
acquired at two successive positions on the path it is possible to estimate the relative movement

SEBD’22: 30th Symposium on Advanced Database Systems, June 19–22, 2022, Tirrenia (Pisa), Italy
⋆
  This research has been made in the context of the Excellence Chair in Computer Engineering at LORIA, University
  of Lorraine, Nancy, France.
*
  Corresponding author.
$ alfredo.cuzzocrea@unical.it (A. Cuzzocrea); majid.abbasi@unical.it (M. A. Sisara); klenac@riteh.hr (K. Lenac);
mumolo@units.it (E. Mumolo)
                                       © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
    CEUR
    Workshop
    Proceedings
                  http://ceur-ws.org
                  ISSN 1613-0073
                                       CEUR Workshop Proceedings (CEUR-WS.org)
�of the object between these two positions. It should be noted that the proposed approach strictly
follows a recent methodology appeared in literature, where artificial intelligence paradigms
marry data analytics methods (e.g., [1, 19, 2]).
   In this paper we describe a family of scan-matching based registration algorithms called
HGLASM-g which perform scan-matching based on a hybrid approach. First, an approximate
pre-alignment of two adjacent maps is performed via a new genetic optimization method called
GLASM-g; then a variant of the Iterative Closest Point (ICP) algorithm is applied to pre-aligned
maps to obtain the final overlap.
   In other proposed genetic scan-matching pre-alignment algorithms, the fitness functions are
based on metrics between actual and reference scan points that require to know the correspon-
dence of point pairs and the translation and rotation between the two scans. However, when
scan acquisitions include noise, correspondence errors may arise. Moreover, also translation
and rotation corrections can lead to errors when they are too large.
   In order to overcome such issues, in this paper we propose a novel metric which does not
require neither points pair correspondences nor translations and rotation corrections. Indeed,
our metric is based on lookup tables built around the reference scan points. The fitness function
weights the hits of actual scan points in the lookup table. The genetic pre-alignment then finds
the scan with the highest fitness within a search space of given size. This guarantees also the
maximum robustness towards both the acquisition errors and the Initial Position Errors (IPE). It
is well known that ICP performance depends on the quality of points pair correspondence and
on the accuracy of the starting point estimation. We overcome this limitation by choosing the
initial guess of ICP via genetic pre-alignment, which makes it close to the true solution. This
way, point correspondence, translation and rotation estimations are performed correctly and,
as a consequence, iteration failures are reduced.
   The algorithms described in this paper form a family in the sense that each algorithm is
characterized by different values of the target search space size. Each size allows us to solve
different registration problems and hence different mobile object tracking scenarios. If the
search space size is small, in fact, the algorithm can recover from small errors only, while, if the
search space is higher, also higher errors can be recovered. However, computation complexity
increases as the search space size gets higher.
   A similar hybrid algorithm is described by Martinez et al. in [17]. Therefore, we consider
the latter algorithm in a comparative approach, and we experimentally show that all the terms
of comparison with this algorithm are improved thanks to the hybrid algorithms proposed
in this paper. Furthermore, we also show that our approach is able to recover from greater
initial positioning and acquisition errors. The key for improvement is the definition of a new
metric used for computing the fitness function of the genetic procedure. The proposed target
scan-matching algorithm is described for the 2D case, but it can be used in the 3D case as well.
Improvements obtained with our proposed algorithm are measured both in terms of accuracy
and noise robustness. Indeed, the estimation of the initial position of target mobile object often
comprises significant errors. For instance, when the mobile object is equipped with a legged or
wheeled locomotion, and the initial position is estimated by means of odometric approaches,
there may be slippage with respect to the floor, which entails significant errors in the initial
position of the object. As a consequence, accuracy of algorithms is seriously affected by IPE.
�2. A Novel Family of Enhanced Hybrid Scan Matching
   Algorithms
The pre-alignment step is inspired by the algorithm called Genetic Lookup based Algorithm for
Scan Matching (GLASM) described by Lenac et al. in [14]. The algorithm described in [14] uses
a metric based on a binary lookup table.




                       A
                (reference scan)    GLASM-g                         MBICP
                       B
                 (new scan with
               odometry estimate)                 x',y',Φ'                        x,y,Φ



Figure 1: Block diagram of the proposed technique.


   In the proposed approach we first improve GLASM by using the 8-bit encoding to store
the probability density of measurements being close to points of reference scan in the lookup
table. We call this improved variant of the existing binary GLASM technique as GLASM-g.
Then, this improved variant of the existing GLASM technique is combined with MbICP [18]. In
more details, Figure 1 shows a block diagram of the proposed approach. Here, the new scan
𝐵, evaluated starting from the odometric estimation of robot movements, is fed as input to
GLASM-g, together with the previous scan, 𝐴. The output of GLASM-g 𝑥′ , 𝑦 ′ , 𝜑′ , is then used as
starting point of the MbICP algorithm that compares 𝐴 and 𝐵, thus producing the final output
𝑥, 𝑦, 𝜑.




Figure 2: A detail of a lookup table surrounding an isolated point of the reference scan. Left: Radial,
Gradual. Right: Squared, Binary.


   In Figure 2 the difference between the two algorithms is shown: in the GLASM-g version the
lookup table is displayed as a gray scale image, while in the GLASM version the lookup table is
seen as a black and white image.
   While in the GLASM lookup table each cell was either 1 or 0 the GLASM-g table can contain
a range of values that model the probability of matching a point using a normal distribution. In
the proposed implementation 256 different values are used requiring 1 byte of memory per cell.
   The computation of the fitness function works as follows. For each point of the new scan
a roto-translation and a discretization are performed to bring the point in the same reference
�frame of the lookup table and select the corresponding cell. However, once the cell is selected,
instead of simply incrementing the fitness with a binary value, the fitness is incremented with a
value corresponding to the probability of matching a point that was saved in the lookup table.
A representation of this probability is shown in Figure 2, left panel, with gray levels.
   In Algorithm 1 we report the pseudo code of the fitness computation used in GLASM-g.

Algorithm 1 GLASM-g Fitness Computation
  Input: 𝐵 // new scan
  Output: 𝑓 𝑖𝑡𝑛𝑒𝑠𝑠
  𝑓 𝑖𝑡𝑛𝑒𝑠𝑠 = 0;
  for (each point 𝑝 of 𝐵) do
      // roto-translation to lookup reference frame
      𝑝′ = changereferenceframe(𝑝)
      𝑓 𝑖𝑡𝑛𝑒𝑠𝑠 = 𝑓 𝑖𝑡𝑛𝑒𝑠𝑠 · 𝑙𝑜𝑜𝑘𝑢𝑝(𝑝′ );
  return 𝑓 𝑖𝑡𝑛𝑒𝑠𝑠;

   The fitness function is essential for the proper functioning of genetic optimization algorithms.
For each individual and for each execution it is computed only once. Its correct definition is
therefore fundamental also from the point of view of computational complexity, given that a
fitness function that requires simple calculations translates into a fast execution of the entire
algorithm, leading to an exploration of a greater search space with a greater success ratio.




Figure 3: Search space size of the algorithms 𝑠𝑚𝑎𝑙𝑙 (A), 𝑚𝑒𝑑𝑖𝑢𝑚 (B), and 𝑙𝑎𝑟𝑔𝑒 (C).


   The goal of two-dimensional scan matching algorithms is to obtain the path of a mobile
object by estimating its relative movements during the path. In other words, we can speak of a
space (𝑡𝑟𝑎𝑛𝑠𝑙𝑎𝑡𝑖𝑜𝑛𝑋 × 𝑡𝑟𝑎𝑛𝑠𝑙𝑎𝑡𝑖𝑜𝑛𝑌 , 𝑟𝑜𝑡𝑎𝑡𝑖𝑜𝑛) within which the scan matching algorithm
searches for its estimate.
   The individual algorithms withing the family differ in the search space size and the parameters
selected for the genetic algorithm. Typically, in order to cover a larger search space, a larger
population is used, as well as the number of generations and runs of the genetic algorithm.
   In this paper we selected and studied three different algorithms from within the family that
we simply call 𝑠𝑚𝑎𝑙𝑙, 𝑚𝑒𝑑𝑖𝑢𝑚, and 𝑙𝑎𝑟𝑔𝑒. The search space size and the combination of genetic
parameters of the selected algorithms are depicted in Figure 3 and listed in Table 1.
�Table 1
Search space size of the selected algorithms and the corresponding parameters used for the genetic
algorithm
                Algorithm         Search Space Size         Genetic Configuration
                                     (𝑑𝑋, 𝑑𝑌, 𝑑𝑅𝑜𝑡)           (𝑝𝑜𝑝 × 𝑔𝑒𝑛 × 𝑟𝑢𝑛𝑠)
                 𝑠𝑚𝑎𝑙𝑙 (A)     ±0.3𝑚, ±0.3𝑚,±17.2∘                20 × 6 × 1
                𝑚𝑒𝑑𝑖𝑢𝑚 (B)     ±1.0𝑚, ±1.0𝑚, ±57.3∘              100 × 10 × 1
                 𝑙𝑎𝑟𝑔𝑒 (C)     ±2.0𝑚, ±2.0𝑚, ±180.0∘             200 × 12 × 2


   The search space of the 𝑠𝑚𝑎𝑙𝑙 algorithm is sufficient to recover from small initial errors,
while 𝑚𝑒𝑑𝑖𝑢𝑚 and 𝑙𝑎𝑟𝑔𝑒 are able to correct progressively larger errors, at the cost of computa-
tional time. The 𝑙𝑎𝑟𝑔𝑒 algorithm is able to recover from arbitrary orientation errors and large
translation errors. The search space is obviously centered on the reference scan.
   For the refinement step we selected the MbICP algorithm [18] which determines the corre-
spondences between the points of the current scan and the reference scan considering both the
rotation and the distance between the points. The algorithm has characteristics of accuracy
and low calculation times. MbICP has better characteristics than other iterative scan matching
algorithms based on point correspondences, for example [21, 23].
   Visual examples of the algorithm in operation are shown in Figure 4. In addition a video file
is available attached to this manuscript that illustrates the problem and shows in more detail
the proposed algorithms in in operation.




Figure 4: Visualization of HGLASM-g operation. In the top-left frame the initial position estimate
of the new scan is depicted in green and the reference scan in blue. The next three frames show the
pre-alignment process depicting the population after 0,5, and 12 generations of the genetic algorithm
with the best estimate shown in red. The last two frames show the refinement of the solution using the
MbICP algorithm (the first and the last iteration are shown).
�3. Experimental Study
The performance of the family of algorithms described in this paper has been verified through
several experiments carried out with a publicly available scan dataset. The dataset used in our
experiments was produced by the RAWSEED project, [4, 6]. The aim of the RAWSEED project
was to produce tools to compare the performance of robotic systems. The laser data considered
in this dataset was obtained from a Sick sensor. To simplify the experiments, a significant subset
of the whole data was randomly extracted across the entire dataset.
   The dataset concerns internal and external environments. The movable objects traced in the
dataset entered and exited from various environments, and also covered urban environments.
The dataset thus provides the opportunity to study the performance of complex navigation
algorithms in various types of environments.
   It is worth to note moreover that the results reported in this Section have been obtained with
a PC equipped with Intel Core 2 Quad Q9550 CPU running at 2,83 GHz.
   The assessment of the accuracy obtained by the algorithms requires not only the laser
scanning data but also the ground truth data. The Rawseed dataset contains ground truth data,
but not for all scans. Furthermore, the reported positioning error is 1 cm, which is not sufficient
for our evaluations. To address this inadequacy, the ground truth values were constructed by
using a translated and rotated copy of the reference scan as the new scan. The ground truth
value is thus known exactly as it corresponds to the position of the reference scan. To prevent
the scan from being compared to a copy of itself the scan was split between the two scans with
the reference scan consisting of the odd readings in the scan, and the new scan consisting of
the even ones.
   In all experiments it is important to study how algorithms converge under different conditions.
To quantitatively express the robustness of the algorithms with respect to initialization errors,
scanning noise and falling in local minima, we defined the Success Ratio. Success Ratio is
the percentage of successful comparisons, computed as the number of successful comparisons
divided by the total number of comparisons. A success is the case in which the resulting
translation and rotation fall within a fixed size ellipsoid around the true values. In all the
experiments the pre-established size is 10 cm for distance and 0.57∘ (0.01[𝑟𝑎𝑑]) for rotation,
respectively. In cases where genetic optimization was used to achieve a coarse pre-alignment of
the hybrid algorithm, the predetermined size was 30 cm and 5.73∘ (0.1[𝑟𝑎𝑑]) respectively. This
relaxation of the thresholds was introduced because the purpose of the pre-alignment is not to
provide an accurate estimate, but an estimate that is close enough to the true position to be
significant for the next step of the algorithm. When the scan matching reached convergence,
the mean and standard deviation of the error (estimated position - true position) were computed
for the considered algorithms.
   In the first set of experiments, the MbICP algorithm was studied. In addition to the obtainable
accuracy, it is important to study how robust is the MbICP to the initial position estimate errors,
i.e. to establish its convergence area. Since the MbICP algorithm is used in the refinement step
of the proposed hybrid approach the pre-alignment step should always produce a good enough
alignment that falls inside this area thus enabling MbICP to further refine the solution.
   In the second set of experiments we compared the genetic algorithm using three different
definitions of the fitness function. The results are described in the 3.1 Section. The results
�presented are fundamental to the algorithm described in this paper.
   Finally, the third set of experiments concerns the comparison of the algorithms proposed in
this paper with the algorithm proposed by Almeida et al. (MbICP) [18] that we have taken as a
reference. These comparisons are listed in the 3.2 Section.

3.1. Genetic Pre-Alignment: Performance Analysis
The performance of the genetic algorithm has been studied by varying the configuration of
genetic parameters such as the size of the population and the number of generations. Figure 5
shows the translation (left) and rotation (right) errors versus the number of evaluations of the
fitness function for the Polar Genetic Algorithm (PGA), GLASM and GLASM-g.




Figure 5: Translation Error (Left) and Rotation Error (Right) Versus the Number of Fitness Evaluations
(Right).


   The calculation time of the algorithms based on the GLASM and GLASM-g lookup tables
respectively represents the time taken only by the matching of the new scan with respect to
the reference scan and not by the initialization phase of the lookup table. The figures show the
times in milliseconds relating to various Success Ratios for 𝑠𝑚𝑎𝑙𝑙 and 𝑙𝑎𝑟𝑔𝑒 search spaces.




Figure 6: Time Needed to Get Various Levels of Success Ratio for Different Sets of Genetic Algorithms
with Small (Left) and Large (Right) Search Size.


  As we can see in Figure 6, the best Success Ratio results, calculation time and accuracy are
always obtained with the fitness function used by the GLASM-g algorithm. In particular we see
�that the GLASM-g algorithm always offers better Success Ratio and calculation speed results
than the PGA algorithm of Martinez et al.. The GLASM algorithm provides the same results as
GLASM-g if the size of the population and the number of generations are both high.
   This slower convergence toward high Success Ratios is explained by the nature of the GLASM
fitness function. Actually, in addition to the smaller marking pattern, the GLASM fitness
function also produces variations of the fitness score that are less refined than GLASM-g.

3.2. Proposed Hybrid Algorithms: Performance Analysis
The objectives of the hybrid algorithm proposed in this paper is to improve the robustness of
the approaches based on Iterative Correspondence Point with respect to the initial positioning
errors, to convergence issues and to the presence of noise in the scans. The increase in robustness
must maintain accuracy and must not be introduced at the expense of computing time. Here we
report the results of the third set of experiments, i.e. those obtained with the hybrid approach
presented in this paper. The parameters of the genetic algorithm are described in the previous
section and shown in Table 1 for the search spaces 𝑠𝑚𝑎𝑙𝑙, 𝑚𝑒𝑑𝑖𝑢𝑚 and 𝑙𝑎𝑟𝑔𝑒.
   We describe now some results concerning the Success Ratios obtained with different values of
initial position errors, both rotation and translation errors. On the other hand, the computation
time slightly increases going from 𝑠𝑚𝑎𝑙𝑙 to 𝑙𝑎𝑟𝑔𝑒 sizes. But it is very important to observe that
also the range of errors that can be corrected increases going from 𝑠𝑚𝑎𝑙𝑙 to 𝑙𝑎𝑟𝑔𝑒 sizes.
   Figure 7 shows the Success Ratio curves versus different values of rotation and translation
initial positioning errors, for 𝑙𝑎𝑟𝑔𝑒 (left) and 𝑚𝑒𝑑𝑖𝑢𝑚 (right) search spaces. In addition to this,
Figure 8 (left) shows the same experimental patterns for 𝑠𝑚𝑎𝑙𝑙 search space.

                   HGLASM-g large                                 HGLASM-g medium




Figure 7: Success Ratio Behavior Versus Rotation and Translation Errors for large (Left) and medium
(Right) Search Space Size.


   As we can see in these figures, the Success Ratio for 𝑙𝑎𝑟𝑔𝑒 search space size is about 90%
for arbitrarily large rotation errors and all considered translation errors up to 2.12 meters.
This large range of errors that can be corrected means that global positioning and tracking
applications are possible. As the search space size decreases the range of errors that can be
corrected is reduced, because the solution is searched in a reduced space. This allows to reduce
the computation time in applications where a smaller search space size is sufficient.
   For comparison we report in the following the Success Ratios obtained with the MbICP
algorithm. Figure 8 (right) shows that the rotation and translation errors that can be corrected
�by the MbICP algorithm are much worse than that corrected by the hybrid algorithm described
in this paper. Just to highlight a couple of results reported in Figure 8 (right), consider the
following initial positioning errors: 20 degree of rotation error and 0.57 meter of translation
error. With these errors we have a Success Ratio of about 93% for large and medium search
space sizes and about 85% for small search space size.
   In the same condition the Success Ratio of the MbICP algorithm drops at about 20%. It is also
important to consider the rotation error after which it is not possible to get significant Success
Ratio. While HGLASM-g medium cannot get significant Success Ratio after 110 degrees and
HGLASM-g small after 60 degree, the MbICP algorithm stops at 40 degrees.
   If the IPE is inside the realignment margins of the local scan matcher, the addition of the
pre-alignment step in theory would not be necessary since the ICP based algorithms is capable
by itself of successful matching. On the first thought the addition of the pre-alignment step
would then just add to execution time incurring speed penalty. However, in practice this often
is not true because the addition of the pre-alignment step typically reduces the number of
iterations for the ICP step. The reason for this speed-up lies in the iterative nature of ICP based
algorithms. If the iteration starts from an initial point that is very far from the true position,
it normally requires many iterations to improve the position. If we look at the last row of the
table we see that if the search space is large, the proposed algorithm obtains a high Success
Ratio in a very short time even if the initial position error is high.
   An important aspect of the proposed algorithms is their robustness against uncertainty. Scan
matching algorithms in general cannot adequately overcome singularity cases like navigation
in long hallways. The uncertainty alongside the direction of the corridor cannot be addressed
using the information contained in the scan alone. The proposed approach however manages to
correct the error along the direction orthogonal to the direction of the hallways even for large
errors of the initial position.

                  HGLASM-g small                                     MBICP




Figure 8: Success Ratio Behavior Versus Rotation and Translation Errors for small Search Space Size
(Left) and with respect to the MbICP algorithm (Right).




4. Conclusions and Future Work
In this paper we have presented a hybrid algorithm for the problem of scan matching. The
algorithm was tested with a dataset consisting of laser scans obtained in various environments.
�The main contribution is that better results of the scan matching operation are obtained in
terms of accuracy and robustness. The main reason is due to the new metric adopted that allows
to compare not just single points but an entire scan. This permits to avoid the preliminary steps
of point correspondence and the translation and rotation computation, that introduce errors in
classical scan matching algorithms.
   Future works will be directed towards the extension of the described hybrid algorithms to the
3-D case, and higher dimensions. Also, we plan to make our comprehensive framework suitable
to the emerging big data trend (e.g., [10, 7, 9, 5, 3]), as to make it able of dealing with specific
features of such innovative settings, like also dictated by some recent studies (e.g., [13, 15, 11]).


Acknowledgments
This research has been partially supported by the French PIA project “Lorraine Université
d’Excellence", reference ANR-15-IDEX-04-LUE.


References
 [1] S, Ahn, S.V. Couture, A. Cuzzocrea, K. Dam, G.M, Grasso, C.K. Leung, K.L. McCormick,
     B.H. Wodi, “A Fuzzy Logic Based Machine Learning Tool for Supporting Big Data Business
     Analytics in Complex Artificial Intelligence Environments,” in Proceedings of FUZZ-IEEE
     2019, 2019, pp. 1–6
 [2] A.-R.A. Audu, A. Cuzzocrea, C.K. Leung, K.A. MacLeod, N.I. Ohin, N.C. Pulgar-Vidal,
     “An Intelligent Predictive Analytics System for Transportation Analytics on Open Data
     Towards the Development of a Smart City,” in Proceedings of CISIS 2019, 2019, pp. 224–236
 [3] L. Bellatreche, A. Cuzzocrea, S. Benkrid, “F&A: A Methodology for Effectively and Effi-
     ciently Designing Parallel Relational Data Warehouses on Heterogenous Database Clusters,”
     in Proceedings of DaWak 2010, 2010, pp. 89–104
 [4] A. Bonarini, W. Burgard, G. Fontana, M. Matteucci, D. G. Sorrenti and J. D. Tardós,
     “RAWSEEDS: Robotics advancement through web-publishing of sensorial and elaborated
     extensive data sets”, in Proceedings of IEEE/RSJ Int. Conf. Intell. Robots Syst. Workshop
     Benchmarks in Robot. Res., 2006, p. 5
 [5] M. Ceci, A. Cuzzocrea, D. Malerba, “Effectively and efficiently supporting roll-up and
     drill-down OLAP operations over continuous dimensions via hierarchical clustering,” in J.
     Intell. Inf. Syst. 44(3), 2015, pp. 309–333
 [6] S. Ceriani, G. Fontana, A. Giusti, D. Marzorati, M. Matteucci, D. Migliore, D. Rizzi, D.
     G. Sorrenti, and P. Taddei, “Rawseeds ground truth collection systems for indoor self-
     localization and mapping,” in Autonomous Robots, vol. 27, no. 4, 2009, pp. 353–371
 [7] G. Chatzimilioudis, A. Cuzzocrea, D. Gunopulos, N. Mamoulis, “A novel distributed frame-
     work for optimizing query routing trees in wireless sensor networks via optimal operator
     placement,” in J. Comput. Syst. Sci. 79(3), 2013, pp. 349–368
 [8] C.-C. Chen, M.-H. Hung, B. Suryajaya, Y.-C. Lin, H. C. Yang, H.-C. Huang, F.-T. Cheng, “A
     Novel Efficient Big Data Processing Scheme for Feature Extraction in Electrical Discharge
     Machining,” in IEEE Robotics Autom. Lett. 4(2), 2019, pp. 910–917
� [9] A. Cuzzocrea, “Combining multidimensional user models and knowledge representation
     and management techniques for making web services knowledge-aware,” in Web Intell.
     Agent Syst. 4(3), 2006, pp. 289–312
[10] A. Cuzzocrea, C. De Maio, G. Fenza, V. Loia, and M. Parente, “OLAP analysis of multidi-
     mensional tweet streams for supporting advanced analytics,” in Proceedings of ACM SAC
     2016 International Conference, 2016, pp. 992–999
[11] J.-E. Deschaud, “IMLS-SLAM: Scan-to-Model Matching Based on 3D Data,” in Proceedings
     of ICRA 2018, 2018, pp. 2480–2485
[12] J. Huang, D. Zhu, Y. Tang, “Health diagnosis robot based on healthcare big data and fuzzy
     matching,” in J. Intell. Fuzzy Syst. 33(5), 2017, pp. 2961–2970
[13] Z. Jiang, J. Zhu, Z. Lin, Z. Li, R. Guo, “3D mapping of outdoor environments by scan
     matching and motion averaging,” in Neurocomputing 372, 2020, pp. 17–32
[14] K. Lenac, E. Mumolo, and M. Nolich, “Robust and accurate genetic scan matching algorithm
     for robotic navigation,” in Lecture Notes in Computer Science, vol. 7101, 2011, pp. 584–593
[15] X. Li, S. Du, G. Li, H. Li, “Integrate Point-Cloud Segmentation with 3D LiDAR Scan-
     Matching for Mobile Robot Localization and Mapping,” in Sensors 20(1), 2020, p. 237
[16] J. Li, R. Zhong, Q. Hu, M. Ai, “Feature-Based Laser Scan Matching and Its Application for
     Indoor Mapping,” in Sensors 16(8), 2016, p. 1265
[17] J.L. Martínez, J. González, J. Morales, A. Mandow, and A.J. García-Cerezo, “Mobile robot
     motion estimation by 2d scan matching with genetic and iterative closest point algorithms,”
     in Journal of Field Robotics, vol. 23, no. 1, 2006, pp. 21–34
[18] J. Minguez, L. Montesano, and F. Lamiraux, “Metric-based iterative closest point scan
     matching for sensor displacement estimation,” in IEEE Trans. Robotics, vol. 22, no. 5, 2006,
     pp. 1047–1054
[19] K.J. Morris, S.D. Egan, J.L. Linsangan, C.K. Leung, A. Cuzzocrea, C.S.H. Hoi, “Token-Based
     Adaptive Time-Series Prediction by Ensembling Linear and Non-linear Estimators: A
     Machine Learning Approach for Predictive Analytics on big Stock Data,” in Proceedings of
     ICMLA 2018, 2018, pp. 1486–1491
[20] X. Niu, T. Yu, J. Tang, L. Chang, “An Online Solution of LiDAR Scan Matching Aided
     Inertial Navigation System for Indoor Mobile Mapping,” in Mob. Inf. Syst. 2017, 2017, pp.
     4802159:1–4802159:11
[21] F. Pomerleau, F. Colas, R. Siegwart, and S. Magnenat, “Comparing ICP variants on real-
     world data sets - open-source library and experimental protocol,” in Auton. Robots, vol. 34,
     no. 3, 2013, pp. 133–148
[22] C. Qian, H. Zhang, J. Tang, B. Li, H. Liu, “An Orthogonal Weighted Occupancy Likelihood
     Map with IMU-Aided Laser Scan Matching for 2D Indoor Mapping,” in Sensors 19(7), 2019,
     p. 1742
[23] S. Rusinkiewicz and M. Levoy, “Efficient variants of the ICP algorithm,” in Proceedings of
     3rd International Conference on 3D Digital Imaging and Modeling (3DIM 2001), 2001, pp.
     145–152
[24] Y. Wu, Q. Su, W. Ma, S. Liu, Q. Miao, “Learning Robust Feature Descriptor for Image
     Registration With Genetic Programming,” in IEEE Access 8, 2020, pp. 39389–39402
[25] D. Zhu, “IOT and big data based cooperative logistical delivery scheduling method and
     cloud robot system,” in Future Gener. Comput. Syst. 86, 2018, pp. 709–715
�