1. Introduction
Positioning in Global Navigation Satellite Systems (GNSS)-denied/challenged indoor/outdoor and transitional environments remains one of the most challenging problems due to the presence of various objects that generally reflect and disperse signals used for positioning, and thus affect the network geometry and data availability. Therefore, such environments and situations require new robust approaches for positioning that mitigate some of these challenges. For that purpose, in this paper we deploy an indoor cooperative positioning (CP) system and test its performance.
CP has demonstrated to be useful for positioning of mobile platforms navigating in challenging GNSS, as well as GNSS-denied environments ([
1,
2,
3,
4]), where at least some of the nodes can achieve an acceptable level of positioning accuracy using GNSS, while other nodes are in GNSS-denied environment(s), and the nodes in the network are well-connected to each other and inter-node range information is available ([
1,
2,
3,
5]). The CP approach relies on information exchange in an inter-connected network of multiple nodes that could be static (called anchor/infrastructure nodes) or dynamic (such as Unmanned Aerial Vehicles (UAVs), pedestrians, cars) in nature ([
1,
2,
3,
4,
6,
7,
8,
9,
10,
11,
12,
13]). Each node shares information about its own state, as well as relative information with respect to its neighbouring nodes. The absolute and relative information received from nodes can be processed to precisely estimate the state (including localisation) of each node. Various CP systems-based on sensors such as Ultra-Wide Band (UWB) [
7], Wireless Fidelity (Wi-Fi) ([
7,
8]), Bluetooth, or other similar sensors ([
9,
10,
11,
12,
13]) have been developed in the literature. These CP systems can be classified according to the type of sensor observations (such as angle of arrival, or return time trip), type of processing architecture used (centralised vs. distributed) and the presence or absence of static anchors (infrastructure-based vs. anchor-free). Compared to distributed architectures, centralised ones offer improved accuracy but at the cost of increased communication and processing requirements. In contrast, distributed architecture offers robustness, scalability, and better reliability of the cooperative network ([
2,
9]). In addition, distributed architectures are adaptive—that is, they can take care of addition or loss of nodes from a network. However, one of the major limitations of distributed algorithms is the presence of unknown correlation among the states of the nodes [
14]. Various distributed algorithms, such as Belief Propagation (BP) ([
9,
10,
11]) and the Covariance Intersection Filter ([
14,
15]) have been proposed for CP. Inclusion of a static anchor (i.e., infrastructure) nodes has been shown to improve localisation accuracy [
2]. On the other hand, anchor-free (or P2P) cooperative systems do not rely on the presence of a fixed infrastructure and can use ad hoc networks for positioning.
In completely GNSS-denied environments, such as an indoor environment, a CP network can be best utilised by realisation of the sufficient number of static anchor nodes, whose precise location is known in advance. Various authors have demonstrated the use of alternative positioning technologies for positioning in GNSS-denied environments (see, e.g., [
6,
7,
8]). In this paper, a CP approach that is reliant on the presence of a large number of static infrastructure/anchor nodes is referred to as peer-to-infrastructure CP (P2I CP), while a peer-to-peer CP (P2P CP) relies only on the communication and information exchange among dynamic nodes. From a practical standpoint, a P2I CP may offer better accuracy as compared to a P2P CP in challenging environments [
16]. However, this is achieved at the cost of increased investment in terms of time and cost required in setting up a large number of static anchor nodes in a P2I framework [
2]. Therefore, a good balance with the two extremes (P2P+P2I) may offer reasonable positioning accuracy, especially in indoor and challenging environments, where GNSS signals may be completely absent.
Recent studies ([
4,
5,
17]) on indoor positioning using UWB demonstrate the usefulness and capabilities of both CP approaches, as well as UWB sensors. The achievable accuracy in indoor environments is reported to be from 30 cm to 2 m. The researchers rely either only on anchor-based positioning [
4] (i.e., P2I) or use a small number of nodes (in some cases, only up to 2) [
17]. In this paper, we aim to show the benefits and capabilities of the centralised P2P+P2I CP under real-world conditions where bad geometry and data availability affect performance. This will be tested on an ad hoc network of 30 infrastructure nodes and for multiple dynamic nodes (i.e., pedestrians). We investigate the use of UWB technologies as an alternative technique for pedestrian CP in an indoor environment. The real-world data used in this paper were collected as part of a benchmarking measurement campaign. The campaign was carried out in the course of the joint International Association of Geodesy (IAG) and the Federation of Surveyors (FIG) working groups’ effort on “Multi-Sensor Systems” at the Ohio State University (OSU) in October 2017 to investigate new approaches and algorithms (more in [
1]). In the experiments (described later), pedestrians jointly navigated together to achieve CP ubiquitous positioning. For the pedestrians, a prototype helmet equipped with a variety of sensors, including two UWB systems, was designed. The results demonstrate that P2P+P2I CP techniques can be useful for the positioning of platforms navigating in swarms or networks and offer significant improvement in terms of positioning accuracy. However, some drawbacks have been observed under good geometry and data availability conditions, where P2I CP performs better. Using the two UWB systems, decimetre-level positioning accuracy is achieved under typical real-world conditions.
The major contribution of this paper is that it experimentally evaluates and compares the performance of P2P and P2P+P2I CP approaches in challenging and constrained indoor environments (such as the hallway), where poor network geometry and poor connectivity among the nodes is frequently observed. Furthermore, this paper also presents a new prototype (currently under development) of a Multi-Sensor Ubiquitous Positioning System (MUPS) specially designed for pedestrians in indoor, outdoor, and transitional environments. This paper not only offers a realistic and practical overview of the performance of UWB-based indoor CP systems, but also demonstrates the existing capabilities of proposed MUPS in real-world environments and highlights that some of these challenges may be mitigated once the proposed MUPS is fully developed and functional.
The remainder of the paper is organised as follows: Firstly, the state-of-the-art UWB sensors and positioning are described in
Section 2, followed by a description of the experimental schemes and test scenarios in combined indoor/outdoor environments, as well as the sensor specifications and characteristics in
Section 3.
Section 4 then covers the positioning framework, and
Section 5 the description and discussion of the main results and findings. Finally,
Section 6 provides a summary and an outlook on future plans for the data-processing and analyses.
4. Positioning Framework
This section presents the mathematical framework used for developing the MUPS. Similarly to [
16], a centralised, constant acceleration, cooperative Extended Kalman Filter (EKF) was implemented. Although this implementation can be generalised to more mobile nodes (as in [
16]), the positioning framework for a network of two mobile nodes and
infrastructure nodes is presented here. The state vectors for two users (i.e., dynamic nodes) are simultaneously estimated. The joint state vector
is given as,
where
and
denote the User’s 1 and 2 state vectors, estimated at the time instant
k. 3D position
, 3D velocity
, and 3D acceleration
were being estimated at each time instant
k for each dynamic node
i where
.
The infrastructure nodes are static, and their positions were determined in the local coordinate system (as shown in
Figure 4). The relative ranges between users enabled P2P CP, and relative ranges from users to infrastructure node enabled P2I communication. The measurement vector is denoted by
and consists of
relative ranges
d between User 1 and infrastructure,
relative ranges
d between User 2 and the infrastructure, and
ranges
d between two users.
The system model (i.e., the evolution of users’ state) is given as,
where
stands for the state evolution function.
is the estimated joint state at the prior time-instant,
. The process noise
is normally distributed with zero mean and covariance matrix
which consists of acceleration noise. The chosen system model is the constant acceleration model due to the use of dynamic nodes (pedestrians in this case) with low and stable accelerations. Given that, the motion model governing the state evolution function
is given as,
The first equation in (
7) denotes the change of position per unit of time—velocity, and the second equation denotes the change of velocity per unit of time—acceleration. As in [
16], joint state evolution can be expressed as,
where the joint transition matrix
is given in (
9) and a state transition matrix
for a single User
i where
, is given in (
10). The general case for
can be found in [
16].
denotes the time increment between two state estimates. The identity matrix is denoted by
I. As seen in [
25], the measurement model can be expressed as,
where
stands for the non-linear measurement model and
stands for measurement noise. In this case, measurement noise was normally distributed with zero mean and variance matrix
. Non-linear measurement equations for user to infrastructure (i.e., P2I) and user to user (i.e., P2P) relative ranges are given below, respectively,
where
indicates the user (i.e., mobile node) in question and
indicates which infrastructure node the measured range refers to. The coordinates of the infrastructure node
j are
. The predicted coordinates of User 1 or 2 at time instant
k are
, where
.
The linearization of the measurement model shown in (
12) and (
13) with respect to any user’s estimated positions, results in,
denotes a Jacobian matrix calculated as a first-order partial derivative of (
12) or (
13) with respect to the User’s
i predicted state
. The simplified joint measurement matrix
is then given as,
The general form of this matrix is given in [
16]. The dimensions of the final joint measurement matrix are given in (
15). The number of rows is equal to the number of available relative ranges at time instant
k, and the number of columns in the matrix corresponds to the total number of states estimated in the joint state vector
. The joint measurement matrix consists of Jacobian matrices calculated for relative ranges to infrastructure nodes
and a Jacobian matrix for the relative range between Users 1 and 2. It is important to note that some matrices in (
15) can be equal to zero. This happens whenever a measurement from a certain infrastructure node is unavailable (i.e., partial derivatives are equal to zero). In the case of P2I positioning only, the Jacobian matrices
and
are equal to zero. Besides the equations specified in this section, the standard EKF equations (as in [
51]) can be used to estimate the position of both users. The following section shows and discusses the results achieved using this approach.
5. Results and Discussion
The goal of this paper was to evaluate the performance of the developed CP system and to draw a conclusion on the benefits of CP capabilities in an indoor environment where the network geometry and data availability can negatively affect performance.
To test this, experiments were performed using the positioning framework (see
Section 4) on three different data sets. Firstly, User Z’s relative ranges to Time Domain infrastructure nodes (i.e., P2I) measured with a Time Domain device set up on a helmet were used. Furthermore, User C’s P2I relative ranges measured with Pozyx UWBs on the helmet to Pozyx infrastructure nodes were used. Finally, a data set made of P2I ranges for both users, with the addition of relative ranges between users (i.e., P2P+P2I) measured with Time Domain devices set up on both users’ helmets, was used. It should be noted here that the User Z was equipped with one UWB device (Time Domain) which measured both P2I and P2P ranges to User C. User C was equipped with two UWB devices, where the Time Domain device measured P2P ranges to User Z and the Pozyx device measured P2I ranges. As shown in
Figure 1, there is a separation of ∼15cm between such two UWB devices, which would affect the CP estimate. However, when estimating the positions, this was not accounted for in this paper, due to the results showing average position error of magnitude ∼15–30 cm under good data availability and geometry conditions. Thus, the effect of different positions of devices was assumed to be small. Data sets used in these experiments were subsets of the collected data described in
Section 3.3. Only part of the full trajectory (as seen in
Figure 5) was used due to the focus of this paper on indoor navigation. The trajectory where users walked through the staircase for the first time to the hallway and back to the staircase (i.e., S1-I1-S2) was chosen. Furthermore, although the framework presented in
Section 4 could be used for 3D positioning, only 2D positions were analysed in this paper, as planimetric accuracy is often more important in flat 2D environments. Additionally, the constraints of flat 2D environments can be imposed on the EKF to improve the accuracy. However, the paper did not impose any such constraints. In future work, when more data are collected on multiple building floors, the performance of the height component will be analysed. A zero-mean Gaussian error distribution with the standard deviation 1 m was assumed for all relative ranges used in the experiments. The acceleration noise was zero-mean Gaussian with the standard deviation of 1 ms
−2. The parameters of EKF (such as
Q and
R) should be determined using a Kalman Filter tuning procedure. However, tuning a centralised EKF can be quite challenging. Therefore, these parameters were determined using a trial and error method. As detailed in
Section 3.3, ground truth checkpoints were used to test MUPS performance in the indoor environment.
Firstly, the performance of User Z was analysed in
Section 5.1 by comparing P2I and P2P+P2I CP solutions with ground truth and between each other. In
Section 5.2, User C’s performance was presented and analysed similarly.
5.1. User Z: P2I vs. P2P+P2I
Figure 9 and
Figure 10 show the estimated positions for User Z for P2I and P2P+P2I CP.
Figure 9 shows an increased position error in the right part of the hallway, which was reduced once the P2P data became available (shown in
Figure 10). Since the hallway is about
m wide, the positioning error is within that value most of the time; except when the position estimate wanders off beyond the walls of the hallway in the right part of the hallway. The accuracy of the position estimate on the staircase seems to be better than in the hallway, as the position estimate is always on the right side of the staircase. Both the better positioning accuracy on the staircase and left side of the hallway may be due to better geometry of nodes. Horizontal Dilution of Precision (HDOP) was used to assess the quality of the geometry. However, due to ground truth data only being available at checkpoints, HDOP (calculated based on ground truth), was also only available then.
Figure 11 and
Figure 12 compare the 2D error with both the HDOP and the number of available measurements for User Z’s P2I and P2P+P2I CP solution, respectively. It should be noted that the comparison of the 2D error with the HDOP is always shown on the upper graph, while the comparison of the 2D error with the number of measurements is always shown in the bottom graph. Furthermore, the estimate of the positioning error is available only when the user was static at the checkpoints (checkpoints are marked with their IDs).
Figure 11 shows that the geometry and availability of measurements are generally good (the lower the HDOP value, the better the geometry). However, an increase of error at checkpoints 1, 34, and 35 coincides with a reduced number of measurements and increased HDOP. The positioning accuracy at checkpoints 36 and 37 is similar to the positioning accuracy at checkpoints 34 and 35, despite having better geometry. As visible in
Figure 11, both geometry and measurement availability improves once the User Z gets to points 36 and 37, however, the accuracy does not improve. This is probably caused by the quality of the measurements User Z receives from anchors on the staircase (anchors 100, 200, and 302) which were not available for checkpoints 34 and 35. The quality of those measurements may be affected due to the signal from nodes 200 and 302 passing twice through the walls for checkpoint 36, and once from nodes 100 and 302, and twice from node 200 for checkpoint 37. Checkpoint 1 has the largest HDOP (i.e., the worst geometry) which may be due to all UWB signals arriving from the same general direction.
Figure 12 shows that the addition of the P2P data did not significantly improve the node geometry, except for checkpoint 1. Although the geometry improved, the error increased there. Most of the time, User Z was receiving the signal from six static nodes where only two of those had direct LOS and all others had to pass through walls two or three times. Thus, the accuracy of positioning was probably affected by the reduced measurement quality.
The statistics for User Z’s CP positioning performance is given in
Table 2. For both data sets (P2I and P2P+P2I), the performance was quantified for every checkpoint. Checkpoints 3–33 show good performance with Root Mean Square Error (RMSE) values ranging from
m to
m. When geometry is bad, as indicated by a high HDOP (shown in
Figure 11), RMSE values range from
m to
m for checkpoints 1, 34, and 35. Although the geometry was good at checkpoints 36 and 37, accuracy was not comparable with the accuracy at points 3–33. As discussed before, this may be a result of the reception of signals from anchor nodes on the staircase, due to which said signals pass through obstacles two or three times.
Table 2 shows that the RMSE at the checkpoints 35 and 36 has been reduced for
and
, respectively, with the addition of the P2P measurements. This does not look to be true for the rest of the checkpoints, where the majority shows to have similar accuracy. However, the RMSE at checkpoint 1 was significantly increased with the addition of P2P ranges, which was not explored further in this paper. The improvement/deterioration of the positioning accuracy shown in
Table 2 is illustrated in
Figure 13.
User Z’s performance shows that the addition of P2P cooperative data to the P2I data can improve positioning accuracy. However, if the performance-based on P2I data only was already good (decimetre level), accuracy remained similar. In some cases, the performance was worse, which may have been due to the introduction of User C’s position uncertainty (bad geometry and low measurement availability).
Section 5.2 provides more detail on this. Thus, a CP system capable of optimising the data used for position estimation at times when P2I data is good enough or P2P is necessary, would probably offer the best solution.
5.2. User C: P2I vs. P2P+P2I
Figure 14 and
Figure 15 show the estimated positions for User C for P2I and P2P+P2I CP. The position estimates of User C-based on P2I data show large errors (see
Figure 14) where position estimates do not only reach outside the limits of the hallway into neighbouring rooms like for User Z.
Figure 14 also shows a zoomed-in view of the estimated positions when only P2I data is used. Position estimates seem to be good on the staircase and at the beginning of the hallway (i.e., the right side of the hallway) where the number of measurements is over three. However, once the user moves on to the hallway where only one measurement is available, the position error significantly increases. This happens due to availability of data from only one infrastructure node Pozyx 682d (see
Figure 16). Thus, the user seems to be walking in a spiral trajectory centred within the UWB anchor Pozyx 682d where the radius of the spiral increases/decreases as the relative range from that infrastructure node increases/decreases. Once the P2P data from User Z is added, positioning errors are reduced (shown in
Figure 15).
It should be noted that the results based on P2I data, in the case of one available range measurement, are not useful since the user could be anywhere in the building. Thus, it is not appropriate to use the proposed framework for positioning based on P2I data only. Still, these results offer an insight into possible issues that can arise due to measurement unavailability and the magnitude of that effect if the system is based on one type of data (relative ranges from anchor nodes in this case). Furthermore, a significant improvement in accuracy was shown once the P2P CP data from only one additional user were fused with P2I data. To make the system more robust, a fusion of P2P+P2I data with IMU data is recommended in the case of unavailability of all relative ranging measurements.
Similarly to
Section 5.1,
Figure 16 and
Figure 17 compare the 2D error with the calculated HDOP and the number of available measurements for User C’s P2I and P2P+P2I CP solution, respectively. Again, positioning error estimation is available only when the user was static at the checkpoints. As shown in
Figure 16, the geometry seems to be good when the User C is on the staircase during the availability of multiple P2I measurements (user is on the staircase before and after he passes checkpoint 686c). The measurement availability and good geometry reflect well on positioning accuracy. The accuracy on the staircase and beginning of the hallway seems to be consistent with the one for User Z under the same conditions. However, the positioning error grows to over 100 m when only one to three P2I measurements are available, due to which the calculated HDOP approaches infinity. Once P2P measurements are added, the geometry improves on the staircase, but not in the hallway where only two to three measurements are still available (see
Figure 17). This metric would improve with the addition of multiple users. Even though the geometry did not improve, communication with User Z offered the needed constraint to the one available P2I measurement in the hallway, which reduced the positioning error from ∼100 m level to mostly meter level.
The statistics for User C’s CP positioning performance is given in
Table 3. The P2I RMSE ranges from
m to
m for checkpoints with good geometry and range availability (staircase and point 37). The points with low measurement availability and thus, very bad geometry show RMSE ranging from
m to
m (points 2–36). Once the P2P data is added, the points with good data availability and geometry have RMSE values ranging from
m to
m, and others from
m to
m. The improvement of accuracy shown in
Table 3 is illustrated in
Figure 18.
In the case of User C, the addition of User Z’s data made a significant impact on the resulting positioning errors for checkpoints 2–36, which were reduced on average by 95%. As for User Z, the RMSE for points that had good measurement availability and geometry was less affected by the addition of P2P ranges. However, some points did show an increase in positioning error. As in the case of User Z, this probably happens due to the addition of position uncertainty of the other user to good P2I data.
These results demonstrate some of the issues for indoor positioning systems. As explained in the earlier sections, this data was collected in a hallway at Ohio State University. The students and staff were not restricted from accessing this area, and that may have affected the quality of the measured ranges in addition to normal UWB radio measurement unavailability, possible multipath from the walls and glass surfaces, and bad geometry between the users and the UWB anchors, resulting in high HDOP values. Thus, these results offer a realistic view into the achievable performance with the presented positioning framework and quantify the effect on the performance some of the possible issues can have in indoor environments (e.g., availability of only one measurement). The benefit of CP in indoor areas where bad geometry and measurement unavailability often happens is demonstrated. Positioning errors may be further reduced with the addition of more users. However, it is also seen that when the infrastructure measurement data is good, P2P data can affect the positioning accuracy negatively. This is one of the drawbacks of CP but could be avoided with network optimisation and possibly with the addition of multiple users. It may be possible to further improve the localisation accuracy by imposing additional constraints, such as information derived from indoor maps. However, this paper assumes that such information is not available to the user, and hence not used in this paper.
6. Conclusions
The data used in this paper were collected as a part of a joint IAG and FIG working groups on a “multi-sensor systems” benchmarking measurement campaign. Although this campaign aimed to explore CP of different platforms (i.e., vehicles, bicyclists, and pedestrians) in GNSS-denied/challenged environments, in this paper, we mainly focused on pedestrian CP navigation in a GNSS-denied environment (i.e., indoors). A specially designed prototype of MUPS was used for data collection. Every prototype was equipped with a variety of sensors. In this paper, we focused on UWB-based CP between two pedestrians navigating jointly within a neighbourhood. An overview of the field experimental schemes, set-ups, characteristics, and sensor specifications, along with the main results for the positioning with two different UWB systems were presented. We demonstrated and discussed the benefits of the MUPS. The test set-ups for the UWB-based CP system in the different test scenarios have proved that they are suitable and practicable. The results show that such a system is able to achieve decimetre-level accuracy for Peer-to-Infrastructure (P2I) CP in conditions of good network geometry and measurement availability. It has to be noted that the measurement collection was carried out during regular office hours where people were moving around freely, which provided realistic performance estimates in regard to good or bad environment conditions. We have also shown that under unfavourable conditions (i.e., bad geometry, low measurement availability), accuracy degrades significantly. If the measurements are available and the network geometry is not good, the experiments show degradation of accuracy from decimetre level to meter level. However, when the measurement availability is very low (i.e., one to two measurements), the accuracy of P2I CP was degraded from decimetre level to average 50-meter level, which has been shown as the biggest disadvantage of the proposed framework due to its inability to constrain one available measurement. That constraint was offered with the additional usage of Peer-to-Peer (P2P) ranges between the two users, which resulted in a significant reduction of positioning error. Had some constraints been introduced to the positioning system (i.e., no running through walls, a pedestrian cannot move 10 m in 1 s), this improvement would not be so severe. However, in the absence of such constraints, the average improvement of does indicate the benefit of CP even with just two users. In cases of good measurement availability and bad network geometry, the addition of P2P to P2I ranges (i.e., P2P+P2I CP) has improved accuracy from meter level to sub-meter levels. When only a few (one or two) measurements were available, positioning accuracy improved from average 50 m level to the average meter level. However, when geometry and measurement availability conditions were good, the addition of P2P CP sometimes negatively affected the performance. Thus, in future, MUPS needs to be capable of network optimisation based on the criteria of geometry and measurement availability. Furthermore, it is expected that even better performance is achievable if communication between multiple dynamic platforms is available. Although the presented CP framework is theoretically scalable to an arbitrary number of nodes, in practice, the network size would be limited by the communication capabilities of the UWB sensors. With the increase in the number of users, we would expect more communication outages among the nodes. However, the proposed EKF is capable of handling large networks in terms of computational requirements.
In future, the integration of other signals-of-opportunities with the existing positioning solution will be tested. In this case, the most prominent signal-of-opportunity, Wi-Fi, will be used for the indoor positioning. We will further focus on analyses of the MUPS positioning performance in the transitional environment (e.g., between outdoor and indoor), which will be enabled by integrating the presented positioning framework with GNSS observations. For that purpose, a modified and extended CP algorithm is currently being developed. Apart from absolute localisation of the users, dead reckoning with the inertial sensors is a goal of future investigations. In addition to the inertial sensors, the use of other smartphone sensors will also be investigated (e.g., the use of a barometric pressure sensor for improving 3D localisation of the users in a multi-storey building). This paper did not consider any network security-related issues in the cooperative positioning system. From a practical perspective, such issues may affect the overall performance and integrity of the system. The network security in cooperative systems will be considered in future research.