**4. Cooperative localization of mobile sensor nodes**

16 Will-be-set-by-IN-TECH

elevation is neglected here as the measurements with a 2-D array do not allow for resolution

The analysis of the corresponding PDPs shows that the impulse responses simulated with the hybrid approach bear much more similarity to the measurements. Although the scatterers are generated in a statistical way, their properties are tightly bound to the properties of underlying reflections so that their contributions do not dominate in the channel impulse response but fill

In the next step, the mean error *μ*e and the standard deviation of the error between the measurement and the hybrid model *σ*<sup>e</sup> is calculated for the path loss *L*, delay spread *σ<sup>D</sup>* and the angular spread *σψ*R. For this purpose, all possible Tx/Rx positions as described in Subsection 3.2 are used. These values and the corresponding values of the error between the

> *μeL* in dB 4.34 1.64 1.98 -0.07 3.27 1.68 *σeL* in dB 0.57 0.77 1.25 1.03 0.73 0.75 *μeσ<sup>D</sup>* in ns 1.70 0.83 1.23 0.04 1.81 -0.76 *σeσ<sup>D</sup>* in ns 0.78 0.59 0.60 0.56 0.66 0.67 *μeψ<sup>R</sup>* in deg 22.35 10.26 19.85 7.15 4.33 -1.19 *σψ<sup>D</sup>* in deg 9.00 9.86 7.10 8.01 5.16 6.03

**Table 1.** Mean values and standard deviations of the error between the measurement and ray tracing

Except for azimuth spread, the standard deviation values are very small. In the case of azimuth spread, however, additional errors are imposed due to path estimation. In a few cases, an insignificant rise is observed. The mean values are improved simultaneously for all

The spread of the error values of path loss, delay spread and capacity resulting from the statistical nature of the model is analyzed also. For this purpose 40 realizations of the channel with the same parameter set are generated. For each realization, the mean error of each channel parameter is calculated. To describe the spread, the standard deviation over all mean

Finally, the derived model is applied to scenario A from Subsection 6.4 to prove the space-time distribution of the additional contributions. The angle dependent PDP simulated with the hybrid method is shown in Fig. 14 . The comparison with Fig. 8 shows that the additional contributions are properly placed in the azimuth-delay space, thus, depicting better the

With this, a simple and effective modeling approach for directional UWB channels is proposed. The ray tracing method is combined with a simple geometric-stochastic model

The parameters of the stochastic model are connected to the properties of reflected paths so that they form a cluster with a certain delay and angle range around the reflected contribution. The stochastic clusters are also implemented around the points of multiple reflections. The

Scenario B Scenario C Scenario D . RT Hyb. RT Hyb. RT Hyb.

the missing dense components of the impulse responses and angular spectra.

measurement and conventional ray tracing are shown in Table 1 .

simulation (RT) and between the measurement and hybrid simulation (Hyb.) .

considered channel characteristics.

clustering effects in the scenario.

which represents the dense part of the channel.

values is adopted.

of paths impinging from below and above the array.

In the application scenario envisaged in the introduction, an unknown environment is inspected by a UWB sensor network. Static anchor nodes of the network are placed at strategic positions. They span a local coordinate system and passively localize people or other moving objects just by electromagnetic waves scattered from them. "Electromagnetic images" of the environment are provided by moving nodes of the network. All data extraction algorithms that evaluate data measured by the sensor network require a priori information about the position of corresponding sensor nodes. In this section, basic principles of the cooperative localization of sensor nodes are described.

UWB localization is usually achieved in two steps, parameter extraction and data fusion, [20, 57]. The parameter extraction estimates parameters of signals received by sensor nodes that are required in the data fusion step. Typical parameters that are used in radio based localization systems are time of arrival (ToA), time difference of arrival (TDoA), angle of arrival (AoA) and/or received signal strength (RSS). The range-based schemes, ToA and TDoA, are shown to yield the best localization accuracy due to the excellent time resolution of UWB signals [19]. The range based ToA approach appears to be the most suitable approach for localization in UWB sensor networks. However, there are still many challenges in developing a real-time ToA based indoor UWB localization system. Due to the number of error sources, such as thermal noise, multipath propagation, direct path (DP) blockage and DP excess delay, the accuracy of the range estimation may get worse. In indoor environments, it is proven that the major sources of errors are multipath components (MPCs) and the NLOS situation [54, 56] that strongly influence the parameter estimation step - the range estimation. The quality of the range estimation is related to the SNR (or distance between Tx and Rx) and the LOS/NLOS situation. It could be improved if suitable a priori information is available. This information is usually obtained from subsequent location estimations. In what follows, we propose a novel UWB localization approach which does not require such a priori information and, instead, is based on the NLOS identification and mitigation.

### **4.1. UWB localization in realistic environments**

The first step in our approach is high precision ToA estimation. Conventionally, ToA estimation for UWB localization is performed via a correlator or equivalently, via a matched filter (MF) receiver, [19]. However, it is difficult and not practical to implement this estimator since the received waveform with many unknown parameters must be estimated. This is almost impossible especially in realistic indoor scenarios. Another approach is the maximum likelihood (ML) based method for joint estimation of path amplitudes and ToAs described e.g. in [31, 67, 76]. This is, however, a computationally extensive method that is not suitable for real-time operations. Although various low-complexity ranging algorithms exist, their performance is not sufficient for high precision ToA estimation. Examples of low-complexity threshold-based methods such as the peak detection method, the fixed threshold method, or the adaptive threshold approach are given e.g. in [11], [17] and [22]. In these approaches, the received signal is compared to an appropriate threshold *δ*, and the first threshold-exceeding sample index corresponds to the ToA estimate, i.e.,

$$
\hat{t}\_{\text{To}A} = t\_{\text{ll}} \ \ n = \min \left\{ i |z[i] \ge \delta \right\}. \tag{4}
$$

<sup>20</sup> <sup>40</sup> <sup>60</sup> <sup>80</sup> <sup>100</sup> <sup>120</sup> <sup>140</sup> <sup>160</sup> <sup>0</sup>

approaches before. location tracking.

Sample Time [s]

where *M* is the MSE of the estimated range estimates.

approximate Maximum likelihood method.

**4.2. Measurement-based verification**

compared with a number of other approaches.

*H*<sup>0</sup> : *M σ*<sup>2</sup>

*H*<sup>1</sup> : *M* > *σ*<sup>2</sup>

ranging from AML Direct range tracking range modification by NLOS ident. Tracking of modif. ranges

(a) (b) **Figure 15.** Data processing results of scenario with 2 LOS nodes and 2 NLOS nodes. (a) Ranging results by using different approaches for one NLOS channel; (b) the localization results by using different

After the NLOS identification, the location estimation is performed by using the identified LOS nodes only. For the implementation of the location estimation, trilateration systems are widely used. Many range-based location estimation methods with different complexity and restrictions have been proposed in the literature. All of them try to acquire a high precision of the location estimate from the range estimates. Different location estimation algorithms, which aim to find the closest position to the current coordinate of the target node, offer different accuracies and complexities. In [55], performances of a number of location estimation algorithms are compared, such as the least squares method, the Taylor series method and the

In order to verify our localization approach described above, a measurement was performed in a radar laboratory environment. Two Rx antennas were situated in one room, another two antennas were situated in the neighboring room and in the corridor. The Tx antenna was mounted on a positioning unit and moved along a predefined rectangular track. The MPD-based algorithm was used for range estimation. The hypothesis test-based NLOS identification and mitigation algorithm, which compares the MSEs of range estimates with the variance of the LOS range estimates, was used for location estimation. For comparison, the approximately Maximum likelihood method was applied for the location estimation, too.

The ranging results obtained for a sensor network containing one NLOS node is shown in Fig. 15(a). The result of the localization is displayed in Fig. 15(b). Both figures illustrate the feasibility of the proposed localization approach and its better performance in most cases

For both, range tracking and location tracking, the Kalman filter was applied.

R1

Cooperative Localization and Object Recognition in Autonomous UWB Sensor Networks 197

*LOS*, no NLOS node exists,

í100 0 100 200 300 400 500 600

*LOS*, NLOS nodes exist. (6)

R3

R2

use original range estimates use direct range tracking results use NLOS identification use modified ranges

R4

Estimated Distance [cm]

For the high precision ToA estimation we proposed an adaptive threshold-based ToA estimation algorithm, the maximum probability of detection (MPD) method, in [56]. It aims at improving the robustness in multipath and NLOS situations. The main idea is to compare the probabilities for a number of possible peaks in the obtained CIR of being the ToA estimates. The probability that a certain sample, e.g. the *i*th sample, is determined as the ToA estimate when its amplitude, *z* [*i*], is equal to or greater than the threshold and the samples before are smaller than it, i.e.,

$$P\_{\mathbf{d}}(i) = \mathbf{P}(\mathfrak{H}\_{\mathbf{ToA}} = i) = \left[ \prod\_{n=1}^{i-1} \mathbf{P}\left(z\left[n\right] < \delta\right) \right] \cdot \mathbf{P}\left(z\left[i\right] \ge \delta\right) \tag{5}$$

where, *n*ˆ *ToA* denotes the estimated index, and *i* = 1, 2, . . . , *N* are the sample indices. The one which has the highest probability leads to the final ToA estimate.

The next step in our localization approach is the NLOS identification and mitigation. The advantage of this approach is that, if the identification is correct, the accuracy of the localization can be considerably improved. Several attempts to cope with the NLOS identification problem have been proposed, such as, methods based on the sudden decrease of the SNR or on the multipath channel statistics, or method by comparing statistics of the estimated distances with a threshold in [7, 68]. However, these methods usually need to record a history of channel statistics. The advantage of our approach based on a hypothesis test proposed in [54] is that it could also be applied in cases, when the target node is static or within the halting period of a moving node. The algorithm compares the mean squared error (MSE) of the estimated range estimates with known variance of the LOS range estimates. The two hypotheses are:

**Figure 15.** Data processing results of scenario with 2 LOS nodes and 2 NLOS nodes. (a) Ranging results by using different approaches for one NLOS channel; (b) the localization results by using different approaches before. location tracking.

$$\begin{cases} H\_0: M \leqslant \sigma\_{LOS'}^2 & \text{no NLOS node exists,} \\ H\_1: M > \sigma\_{LOS'}^2 & \text{NLOS nodes exist.} \end{cases} \tag{6}$$

where *M* is the MSE of the estimated range estimates.

18 Will-be-set-by-IN-TECH

situation. It could be improved if suitable a priori information is available. This information is usually obtained from subsequent location estimations. In what follows, we propose a novel UWB localization approach which does not require such a priori information and, instead, is

The first step in our approach is high precision ToA estimation. Conventionally, ToA estimation for UWB localization is performed via a correlator or equivalently, via a matched filter (MF) receiver, [19]. However, it is difficult and not practical to implement this estimator since the received waveform with many unknown parameters must be estimated. This is almost impossible especially in realistic indoor scenarios. Another approach is the maximum likelihood (ML) based method for joint estimation of path amplitudes and ToAs described e.g. in [31, 67, 76]. This is, however, a computationally extensive method that is not suitable for real-time operations. Although various low-complexity ranging algorithms exist, their performance is not sufficient for high precision ToA estimation. Examples of low-complexity threshold-based methods such as the peak detection method, the fixed threshold method, or the adaptive threshold approach are given e.g. in [11], [17] and [22]. In these approaches, the received signal is compared to an appropriate threshold *δ*, and the first threshold-exceeding

For the high precision ToA estimation we proposed an adaptive threshold-based ToA estimation algorithm, the maximum probability of detection (MPD) method, in [56]. It aims at improving the robustness in multipath and NLOS situations. The main idea is to compare the probabilities for a number of possible peaks in the obtained CIR of being the ToA estimates. The probability that a certain sample, e.g. the *i*th sample, is determined as the ToA estimate when its amplitude, *z* [*i*], is equal to or greater than the threshold and the samples before are

> *<sup>i</sup>*−<sup>1</sup> ∏*n*=1

where, *n*ˆ *ToA* denotes the estimated index, and *i* = 1, 2, . . . , *N* are the sample indices. The one

The next step in our localization approach is the NLOS identification and mitigation. The advantage of this approach is that, if the identification is correct, the accuracy of the localization can be considerably improved. Several attempts to cope with the NLOS identification problem have been proposed, such as, methods based on the sudden decrease of the SNR or on the multipath channel statistics, or method by comparing statistics of the estimated distances with a threshold in [7, 68]. However, these methods usually need to record a history of channel statistics. The advantage of our approach based on a hypothesis test proposed in [54] is that it could also be applied in cases, when the target node is static or within the halting period of a moving node. The algorithm compares the mean squared error (MSE) of the estimated range estimates with known variance of the LOS range estimates. The

P (*z* [*n*] < *δ*)

· P (*z* [*i*] ≥ *δ*), (5)

<sup>ˆ</sup>*tToA* <sup>=</sup> *tn*, *<sup>n</sup>* <sup>=</sup> min {*i*|*z*[*i*] <sup>≥</sup> *<sup>δ</sup>*} . (4)

based on the NLOS identification and mitigation.

**4.1. UWB localization in realistic environments**

sample index corresponds to the ToA estimate, i.e.,

*P*d(*i*) = P(*n*ˆToA = *i*) =

which has the highest probability leads to the final ToA estimate.

smaller than it, i.e.,

two hypotheses are:

After the NLOS identification, the location estimation is performed by using the identified LOS nodes only. For the implementation of the location estimation, trilateration systems are widely used. Many range-based location estimation methods with different complexity and restrictions have been proposed in the literature. All of them try to acquire a high precision of the location estimate from the range estimates. Different location estimation algorithms, which aim to find the closest position to the current coordinate of the target node, offer different accuracies and complexities. In [55], performances of a number of location estimation algorithms are compared, such as the least squares method, the Taylor series method and the approximate Maximum likelihood method.

### **4.2. Measurement-based verification**

In order to verify our localization approach described above, a measurement was performed in a radar laboratory environment. Two Rx antennas were situated in one room, another two antennas were situated in the neighboring room and in the corridor. The Tx antenna was mounted on a positioning unit and moved along a predefined rectangular track. The MPD-based algorithm was used for range estimation. The hypothesis test-based NLOS identification and mitigation algorithm, which compares the MSEs of range estimates with the variance of the LOS range estimates, was used for location estimation. For comparison, the approximately Maximum likelihood method was applied for the location estimation, too. For both, range tracking and location tracking, the Kalman filter was applied.

The ranging results obtained for a sensor network containing one NLOS node is shown in Fig. 15(a). The result of the localization is displayed in Fig. 15(b). Both figures illustrate the feasibility of the proposed localization approach and its better performance in most cases compared with a number of other approaches.

20 Will-be-set-by-IN-TECH 198 Ultra-Wideband Radio Technologies for Communications, Localization and Sensor Applications Cooperative Localization and Object Recognition in Autonomous UWB Sensor Networks <sup>21</sup>


Track1

**...**

**...**

x (m)

**Figure 17.** UWB MIMO imaging scenario. "Track 1 & 2" are the transmitter tracks; Triangles: nonlinear

the environment. The receivers collect the backscattered probing signals to produce an image of the environment. Meanwhile, they could also serve as anchor sensors to support other applications, such as localizing and tracking the position of the moving transmitters [61].

The sensor motion factor is shown in Fig. 18 (a) and (b), with respect to different tracks (Track1 and Track2, as defined in Fig. 17). In the figures, apparently, the ripples are narrower in the direction of L1 compared to the direction of L2. It indicates that the resolving performance in the direction of L1 is better than that of L2, due to the total angular rotation in the direction of L1 is far greater than the angular rotation in the direction of L2 with respect to the reference **x**0. For similar reasons, the resolving performance of "Track1" is better than the resolving performance of "Track2" in the corresponding directions. In addition, it is shown in Fig. 18 (a) that a "ghost" object occurs in the direction of L2, due to an insufficient illumination of the object. Generally, it would generate a false object image, and consequently worsen the quality

In Fig. 18 (a) and (b), the motion factors are given with respect to linear tracks. However, in practice, the sensors are not necessarily moving along linear tracks. There may be more practical irregular tracks as shown in Fig. 17 where the triangles indicate the transmission positions. The irregular movement of the sensors could improve the performance of ghost suppression, since the irregular tracks can provide a more sufficient illumination of the

According to the sensor topology in Fig. 17, the topology factor is given in Fig. 18(c). In the figure, the ghost image is partially suppressed. As shown in the figure, the suppression residuals exist at the ghost image position. However, they are not as strong as the real object. Theoretically, the ghost image can be further suppressed by optimizing the sensor spatial

Figure 18 (a), (b) and (c) indicate the resolution contribution of the sensor motions and the sensor placement topology to the overall resolution. As given in Fig. 17, the overall performance of the system is the combination of all involved individual factors. It implies that we can try to realize a better overall resolving performance by (i) optimizing each individual factor, or (ii) trading-off between related factors. For example, in order to suppress the "ghost" image, on the one hand, we can optimize the movement tracks via the motion factor and the sensor placement topology via the topology factor. On the other hand, a compromise can be made between the motion factor and the topology factor. In this sense, due to the interaction

Ti

Rj

**... ...**

Cooperative Localization and Object Recognition in Autonomous UWB Sensor Networks 199

Track2

y (m)

environment compared to the linear tracks.

tracks.

of the image.

placement.

x 0

L2 L1

**Figure 16.** Factorization of MIMO ambiguity function and its potential applications.
