**9. Detection and localization of moving objects**

In this section we describe methods for the detection and localization of several moving persons who do not have a tag or device attached to them. This is useful for applications where the targets being tracked are not expected to cooperate with the system.

A single UWB sensor suffers from narrow-band interferences as well as shadowing when detecting multiple targets. Often, the closest target can be observed best. Due to shadowing caused by the closest target, the other targets are usually invisible, although they lie within the coverage of the sensor. The closest target attenuates the electromagnetic waves that propagate toward the targets located behind it. Thus, the shadowed targets are almost impossible to detect by a single sensor node. Using a distributed network of UWB radars, the estimated target positions are refined by fusing the information available from the multiple sensors which are able to detect the targets. It also allows for the observation of the targets from different viewing angles.

−7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8

X [m]

(a)

Staircases Coridoor Coridoor

**Sensor 3 Sensor 2**

Receiver Receiver

Transmitter

**Sensor 6** Windows

**Sensor 1**

**Figure 56.** a) The verification scenario and b) the received impulse responses after background

to the sensor evokes the strongest echo, and almost completely shadows the other two targets when it is very close to the sensor. In order to simplify the localization algorithm, we make the assumption that each sensor node can observe only the closest target. This assumption allows the algorithms for multi-target detection such as the constant false alarm rate (CFAR) techniques [15], algorithms for data association, or multi-hypotheses tests to be skipped. This makes the proposed algorithm computationally efficient and suitable for real-time applications. On the other hand, this simplification restricts the number of targets that can be localized by the approach. A sensor network with *N* sensor nodes can localize up to *N* targets in real time. A more detailed description of this method is presented in [72].

After the time-invariant background signals have been removed from the received impulse response, the targets' ranges need to be estimated. The distance from the transmitter to the target and back to the receiver is considered as target range. The target range of the closest target can be estimated by using one of the various threshold-based approaches where the leading edge of the received signal is detected [10–12, 22, 56]. Threshold-based approaches offer simple techniques which detect the leading edge of a received signal by comparing the received signal magnitude or energy with a predefined threshold. The choice of an

As in our scenario each sensor has two receivers, from each sensor node we have the estimated range of the closest target by the two receivers. The estimated ranges determine two ellipses whose focal points are determined by the locus of the transmitting antenna and the locus of the respective receiving antenna. The target position is determined as the intersection of the two ellipses within the area of detection of the respective sensor. The location estimates with respect to each of the six sensor nodes is shown in Fig. 57(a). Here, the estimates are

appropriate threshold is mandatory for a good performance of this estimator.

represented by the same color as the respective sensor node.

**Sensor 5**

Time of Arrival [ns]

Cooperative Localization and Object Recognition in Autonomous UWB Sensor Networks 231

Measurement time [s]

(b)

5 10 15 20 25 30 35 40 45

subtraction

Y [m]

**Sensor 4**

*9.1.1. Single target range estimation*

*9.1.2. Localization*

A network of multiple static bat-type sensors distributed around the inspected area is used. Each bat-type sensor node is capable of autonomously detecting and localizing the present targets. The weak echoes of the targets are first detected in the backscattered signal, after which the detections are correctly assigned to the targets and finally, the target information from all sensors is fused together resulting in location estimates of the targets in the scenario. Here we describe two methods that can cope with this challenging task. The first one uses simplification assumptions of one target detection per sensor to combat the data association problem. It is described in Section 9.1. The second one is based on the probabilistic hypothesis density (PHD) filter for range estimation and position tracking and is described in more detail in Section 9.2.

In both methods, background subtraction is used for target detection [9, 21, 38, 71]. The echoes evoked by the moving targets are usually weak and must be detected in the presence of other strong multipath signals. The disturbing signals are usually time invariant and overlay echoes from the target. Therefore, one received impulse response is insufficient to separate them. However, due to their time-invariance these background signals can be estimated by a suitable estimation technique from a sequence of received impulse responses. The subtraction of the estimated background from the measured data leads to a signal where the weak target echo can be recognized (see Fig. 56(b)) and its range can be estimated more easily.

For the verification of the methods, a sensor network constellation as in Fig. 56(a) is used. Five UWB bat-type sensors are placed around a large foyer, and one bat-type sensor is placed behind one of the walls, directed towards the area of interest. Directional horn antennas with different size and quality are used, resulting in variable target detection performance between the sensors. A scenario with three moving persons was used. All three persons move in the area of interest during the whole measurement time. They move in a straight line from one wall to the opposite and back. The starting position of the targets is shown by a star in Fig. 56(a). The arrow signifies the direction of movement at the start of the measurement.

## **9.1. Low complexity method for the localization of multiple targets in sensor networks**

The method presented in this section was proposed as a computationally efficient method for the localization of multiple targets in dense sensor networks. Each target can be usually observed by at least one sensor node. As can be seen in Fig. 56(b), the closest target

**Figure 56.** a) The verification scenario and b) the received impulse responses after background subtraction

to the sensor evokes the strongest echo, and almost completely shadows the other two targets when it is very close to the sensor. In order to simplify the localization algorithm, we make the assumption that each sensor node can observe only the closest target. This assumption allows the algorithms for multi-target detection such as the constant false alarm rate (CFAR) techniques [15], algorithms for data association, or multi-hypotheses tests to be skipped. This makes the proposed algorithm computationally efficient and suitable for real-time applications. On the other hand, this simplification restricts the number of targets that can be localized by the approach. A sensor network with *N* sensor nodes can localize up to *N* targets in real time. A more detailed description of this method is presented in [72].

### *9.1.1. Single target range estimation*

After the time-invariant background signals have been removed from the received impulse response, the targets' ranges need to be estimated. The distance from the transmitter to the target and back to the receiver is considered as target range. The target range of the closest target can be estimated by using one of the various threshold-based approaches where the leading edge of the received signal is detected [10–12, 22, 56]. Threshold-based approaches offer simple techniques which detect the leading edge of a received signal by comparing the received signal magnitude or energy with a predefined threshold. The choice of an appropriate threshold is mandatory for a good performance of this estimator.

### *9.1.2. Localization*

52 Will-be-set-by-IN-TECH

In this section we describe methods for the detection and localization of several moving persons who do not have a tag or device attached to them. This is useful for applications

A single UWB sensor suffers from narrow-band interferences as well as shadowing when detecting multiple targets. Often, the closest target can be observed best. Due to shadowing caused by the closest target, the other targets are usually invisible, although they lie within the coverage of the sensor. The closest target attenuates the electromagnetic waves that propagate toward the targets located behind it. Thus, the shadowed targets are almost impossible to detect by a single sensor node. Using a distributed network of UWB radars, the estimated target positions are refined by fusing the information available from the multiple sensors which are able to detect the targets. It also allows for the observation of the targets from

A network of multiple static bat-type sensors distributed around the inspected area is used. Each bat-type sensor node is capable of autonomously detecting and localizing the present targets. The weak echoes of the targets are first detected in the backscattered signal, after which the detections are correctly assigned to the targets and finally, the target information from all sensors is fused together resulting in location estimates of the targets in the scenario. Here we describe two methods that can cope with this challenging task. The first one uses simplification assumptions of one target detection per sensor to combat the data association problem. It is described in Section 9.1. The second one is based on the probabilistic hypothesis density (PHD) filter for range estimation and position tracking and is described in more detail

In both methods, background subtraction is used for target detection [9, 21, 38, 71]. The echoes evoked by the moving targets are usually weak and must be detected in the presence of other strong multipath signals. The disturbing signals are usually time invariant and overlay echoes from the target. Therefore, one received impulse response is insufficient to separate them. However, due to their time-invariance these background signals can be estimated by a suitable estimation technique from a sequence of received impulse responses. The subtraction of the estimated background from the measured data leads to a signal where the weak target echo

For the verification of the methods, a sensor network constellation as in Fig. 56(a) is used. Five UWB bat-type sensors are placed around a large foyer, and one bat-type sensor is placed behind one of the walls, directed towards the area of interest. Directional horn antennas with different size and quality are used, resulting in variable target detection performance between the sensors. A scenario with three moving persons was used. All three persons move in the area of interest during the whole measurement time. They move in a straight line from one wall to the opposite and back. The starting position of the targets is shown by a star in Fig. 56(a). The arrow signifies the direction of movement at the start of the measurement.

**9.1. Low complexity method for the localization of multiple targets in sensor**

The method presented in this section was proposed as a computationally efficient method for the localization of multiple targets in dense sensor networks. Each target can be usually observed by at least one sensor node. As can be seen in Fig. 56(b), the closest target

can be recognized (see Fig. 56(b)) and its range can be estimated more easily.

where the targets being tracked are not expected to cooperate with the system.

**9. Detection and localization of moving objects**

different viewing angles.

in Section 9.2.

**networks**

As in our scenario each sensor has two receivers, from each sensor node we have the estimated range of the closest target by the two receivers. The estimated ranges determine two ellipses whose focal points are determined by the locus of the transmitting antenna and the locus of the respective receiving antenna. The target position is determined as the intersection of the two ellipses within the area of detection of the respective sensor. The location estimates with respect to each of the six sensor nodes is shown in Fig. 57(a). Here, the estimates are represented by the same color as the respective sensor node.

#### 54 Will-be-set-by-IN-TECH 232 Ultra-Wideband Radio Technologies for Communications, Localization and Sensor Applications Cooperative Localization and Object Recognition in Autonomous UWB Sensor Networks <sup>55</sup>

**Figure 57.** The target position estimates a) by each sensor and b) after data fusion

### *9.1.3. Data fusion*

The location estimates from all sensor nodes must be fused together by an appropriate algorithm. We propose an imaging algorithm that does not require any data association. It creates a sampled image of the inspected area at a given time *i* which is stored in a matrix **Pi**. Each element of the matrix corresponds to the spatial coordinates within the inspected area. The image is updated according to

$$\mathbf{P\_i} = \alpha \mathbf{P\_{i-1}} + (1 - \alpha) \mathbf{U\_i} \tag{23}$$

Gaussian mixture (GM) approach. We first use it for estimating the target ranges with respect to each sensor, and later for fusion of the target location estimates by all sensors. The method

Cooperative Localization and Object Recognition in Autonomous UWB Sensor Networks 233

In Fig. 58 we describe the processing done on a measured impulse response from raw measurement to range estimation. The impulse response presented is from a scenario with three moving targets. First the measured impulse response is shown in dark blue. As can be seen, the target echoes are non-detectable. After subtracting the time-invariant background (resulting signal shown in green) the echoes of the moving targets are detectable more easily. A Gaussian adaptive threshold constant false alarm rate (CFAR) method as in [15] is used for extracting the ranges from the resulting signal. The adaptive threshold calculated is shown in cyan. A Neyman-Pearson detector is used to discriminate between the noise and the target echo. The ranges extracted using the CFAR approach are shown as points in magenta.

The CFAR detector is not immune to clutter noise and false detections. As we can see in Fig. 58, four targets are detected by the CFAR method although there are only three targets in the scenario. In addition, there are also multiple detections per target, making it difficult to decide if there are multiple targets in the vicinity of each other, or it is only a single target. By using a GM-PHD filter we improve the target range estimates. The ranges extracted by the CFAR method are used as observations for the PHD filter, and a linear Gaussian model is assumed. The target detection and survival probability are assumed to be state independent. To estimate the number of targets and their states, first the Gaussian terms with low weights are pruned and the Gaussian terms that are within a small distance of each other are merged together. The number of targets is estimated by the number of Gaussian terms with a weight above a predefined threshold, and their state is represented by the mean terms of these Gaussian mixtures. The surviving Gaussian terms for the impulse response given in Fig. 58 are shown

As in Section 9.1.2, the target locations are computed using the ranges estimated by the two receivers of the sensor. As multiple targets are detected by each receiver, the range estimates from both receivers corresponding to the same target need to be associated. An intersection threshold *Ts* is defined for each sensor. The ranges of a target in the inspected area with respect

then defined as the maximum possible absolute difference between these ranges such that the

As the size of the inspected area is known or approximated by the sensor detection range, the intersection threshold is calculated prior to scanning the environment. The range association is done such that for each range estimate from the first receiver, we associate the range estimate

*Ts* = max *<sup>k</sup>*∈*<sup>A</sup>* <sup>|</sup>*<sup>r</sup> s*,1 *<sup>k</sup>* − *r s*,2

> |*r s*,1 *<sup>k</sup>* − *r s*,2

*s*,1 *<sup>k</sup>* and *r*

*s*,2

*<sup>k</sup>* . The intersection threshold is

*<sup>k</sup>* | (24)

*<sup>k</sup>* | < *Ts* (25)

is explained in greater detail in [27].

*9.2.1. Multiple target range estimation*

in black.

*9.2.2. Target localization*

target lies in the area of interest.

from the second receivers which satisfies

to both receivers of a sensor *s* are first calculated as *r*

where **Ui** is the innovation matrix and *α* is the forgetting factor. Thus, the new image estimate **Pi** takes a fraction of the previous estimate **Pi**−**<sup>1</sup>** and a fraction from the innovation matrix **Ui**. The innovation matrix **Ui** maps the location estimates of all sensor nodes. The matrix **Pi** indicates moving targets in the environment as hot spots within the image.

An example of the data fusion is given in Fig. 57(b). It depicts one snapshot of the inspected area computed for a certain measurement time. Due to the exponential averaging each target appears in the snapshot like a "comet" with a tail. This tail indicates previous locations of the objects. Its length can be adjusted by adjusting the weighting factor *α* in Eqn. 23.

### **9.2. Localization and tracking of multiple targets in sensor networks based on the PHD filter**

Unlike the method presented in Section 9.1, this method uses the range information related to all targets detected by each sensor. Traditionally, multiple targets are tracked using two-step approaches: data association of the observations to targets followed by single target tracking on the associated data. The traditional approaches are generally computationally expensive, thus, we use a less computationally expensive alternative based on random finite sets (RFS) [36, 37]. It allows for tracking of time-variant number of unknown moving targets in the presence of false alarms, miss-detections and clutter. An approximation of the Bayesian multi-target tracking represented by RFSs, by its first order moment leads to the PHD filter [27]. There are two implementations of the PHD filter, one based on a sequential Monte Carlo approach [65], and the other based on Gaussian mixtures [66]. Here, we use the Gaussian mixture (GM) approach. We first use it for estimating the target ranges with respect to each sensor, and later for fusion of the target location estimates by all sensors. The method is explained in greater detail in [27].

### *9.2.1. Multiple target range estimation*

54 Will-be-set-by-IN-TECH

Y−direction [m]

The location estimates from all sensor nodes must be fused together by an appropriate algorithm. We propose an imaging algorithm that does not require any data association. It creates a sampled image of the inspected area at a given time *i* which is stored in a matrix **Pi**. Each element of the matrix corresponds to the spatial coordinates within the inspected area.

where **Ui** is the innovation matrix and *α* is the forgetting factor. Thus, the new image estimate **Pi** takes a fraction of the previous estimate **Pi**−**<sup>1</sup>** and a fraction from the innovation matrix **Ui**. The innovation matrix **Ui** maps the location estimates of all sensor nodes. The matrix **Pi**

An example of the data fusion is given in Fig. 57(b). It depicts one snapshot of the inspected area computed for a certain measurement time. Due to the exponential averaging each target appears in the snapshot like a "comet" with a tail. This tail indicates previous locations of the

**9.2. Localization and tracking of multiple targets in sensor networks based on the**

Unlike the method presented in Section 9.1, this method uses the range information related to all targets detected by each sensor. Traditionally, multiple targets are tracked using two-step approaches: data association of the observations to targets followed by single target tracking on the associated data. The traditional approaches are generally computationally expensive, thus, we use a less computationally expensive alternative based on random finite sets (RFS) [36, 37]. It allows for tracking of time-variant number of unknown moving targets in the presence of false alarms, miss-detections and clutter. An approximation of the Bayesian multi-target tracking represented by RFSs, by its first order moment leads to the PHD filter [27]. There are two implementations of the PHD filter, one based on a sequential Monte Carlo approach [65], and the other based on Gaussian mixtures [66]. Here, we use the

X−direction [m]

(b)

**Pi** = *<sup>α</sup>***Pi**−**<sup>1</sup>** + (<sup>1</sup> − *<sup>α</sup>*)**Ui** (23)

−6 −4 −2 0 2 4 6 8

Target 3

Target 2

Receivers

Target 1

Transmitters

Sensor 5

**Figure 57.** The target position estimates a) by each sensor and b) after data fusion

indicates moving targets in the environment as hot spots within the image.

objects. Its length can be adjusted by adjusting the weighting factor *α* in Eqn. 23.

−6 −4 −2 0 2 4 6 8

X−direction [m]

Sensor 4 Sensor 1

(a)

The image is updated according to

Sensor 3 Sensor 6 Sensor 2

*9.1.3. Data fusion*

**PHD filter**

Y−direction

In Fig. 58 we describe the processing done on a measured impulse response from raw measurement to range estimation. The impulse response presented is from a scenario with three moving targets. First the measured impulse response is shown in dark blue. As can be seen, the target echoes are non-detectable. After subtracting the time-invariant background (resulting signal shown in green) the echoes of the moving targets are detectable more easily. A Gaussian adaptive threshold constant false alarm rate (CFAR) method as in [15] is used for extracting the ranges from the resulting signal. The adaptive threshold calculated is shown in cyan. A Neyman-Pearson detector is used to discriminate between the noise and the target echo. The ranges extracted using the CFAR approach are shown as points in magenta.

The CFAR detector is not immune to clutter noise and false detections. As we can see in Fig. 58, four targets are detected by the CFAR method although there are only three targets in the scenario. In addition, there are also multiple detections per target, making it difficult to decide if there are multiple targets in the vicinity of each other, or it is only a single target. By using a GM-PHD filter we improve the target range estimates. The ranges extracted by the CFAR method are used as observations for the PHD filter, and a linear Gaussian model is assumed. The target detection and survival probability are assumed to be state independent. To estimate the number of targets and their states, first the Gaussian terms with low weights are pruned and the Gaussian terms that are within a small distance of each other are merged together. The number of targets is estimated by the number of Gaussian terms with a weight above a predefined threshold, and their state is represented by the mean terms of these Gaussian mixtures. The surviving Gaussian terms for the impulse response given in Fig. 58 are shown in black.

### *9.2.2. Target localization*

As in Section 9.1.2, the target locations are computed using the ranges estimated by the two receivers of the sensor. As multiple targets are detected by each receiver, the range estimates from both receivers corresponding to the same target need to be associated. An intersection threshold *Ts* is defined for each sensor. The ranges of a target in the inspected area with respect to both receivers of a sensor *s* are first calculated as *r s*,1 *<sup>k</sup>* and *r s*,2 *<sup>k</sup>* . The intersection threshold is then defined as the maximum possible absolute difference between these ranges such that the target lies in the area of interest.

$$T\_s = \max\_{k \in A} |r\_k^{s,1} - r\_k^{s,2}| \tag{24}$$

As the size of the inspected area is known or approximated by the sensor detection range, the intersection threshold is calculated prior to scanning the environment. The range association is done such that for each range estimate from the first receiver, we associate the range estimate from the second receivers which satisfies

$$|r\_k^{s,1} - r\_k^{s,2}| < T\_s \tag{25}$$

amount of noise due to small errors in the range estimation. A simplified GM-PHD filter is applied for fusing the target location estimates from all sensors. The location estimates are used to form the observation RFS for the GM-PHD filter. The target state is defined by the 2D

Cooperative Localization and Object Recognition in Autonomous UWB Sensor Networks 235

neighbor association, and the results for the three-target scenario can be seen in Fig. 59(b). Both methods presented in this section can be used for the localization of multiple non-cooperative targets using distributed UWB sensor network with static sensors as in the

The CoLOR project was devoted to the recognition of unknown environments using UWB technology. This topic encompasses a number of partial challenges. In order to obtain a complex picture of some catastrophic scenario, like the detection of victims after some natural disasters, their location in collapsed buildings, the geometrical information and the status of the buildings, we derived new detection, localization and imaging algorithms. Their performance was analyzed on simulated data and data measured in realistic scenarios using UWB sensors. These UWB sensors are capable of real-time operation in MIMO configuration. This allows us to analyze the application of UWB sensor networks and cooperative approaches for the localization of sensor nodes within the network, for the localization of people, for the detection and monitoring of their live signs, and for the imaging

It was shown that by using a mobile UWB radar with multiple antennas, it is possible in an efficient way to reconstruct the basic layout of rooms and the position of freestanding objects. The detected features are added to a map while at the same time the own position is estimated (SLAM). To minimize the computational cost and the number of measurements needed, simplified models for wave propagation and stochastic, dynamic state space estimators where enhanced. The method of data association proved to be most critical regarding the precision

Using this map as a-priori information, the detection, localization and the imaging of the objects within an indoor scenario can be performed using the developed localization and imaging algorithms. By knowing the location of the individual objects, the potential of UWB radar was fully tapped by obtaining super-resoluted local information about 2D as well as 3D complex objects (concerning the outer contour). The interior of objects was gathered by novel algorithms which are based on exact radiation patterns depending on the permittivity of the medium while showing low computational load. The obtained radar images are post-processed by means of object recognition algorithms designed for full, fragmented or restricted illumination to provide recognition of the object under test from a finite alphabet. By the adaption of classical ellipsometry to the UWB-range, an estimation of dielectric surface properties can robustly be performed even for small dimensioned objects with a size of a couple of wavelengths. In addition polarimetric measurements as well as polarimetric data processing were taken into account to obtain object features which may remain invisible in

In order to test and compare different algorithms and antenna arrangements for indoor UWB sensing and imaging, a realistic UWB multi-path propagation simulation tool was developed. The propagation model is based on a hybrid approach which combines the deterministic ray tracing method (based on geometrical optics and the uniform theory of diffraction)

*<sup>T</sup>*. The targets are identified using nearest

target coordinates and velocity vector, **x** = [*x x y*˙ *y*˙]

scenario described in Section 2.1 .

**10. Conclusion**

of their surroundings.

and reliability of the map.

mono polarized systems.

**Figure 58.** Target echo detection - measured impulse response (blue), normalized signal magnitude after background subtraction (green), CFAR test statistic (red), CFAR adaptive threshold (cyan), indices of detected targets by CFAR (magenta) and Gaussian mixtures representing the estimated target ranges (black) are shown

as a range estimate that corresponds to the same target. When multiple range estimates comply with this rule, the range estimate which results in the smallest absolute difference is chosen.

The target location is analytically calculated as the intersection of the ellipses defined by the associated range estimates. The location estimates with respect to each of the six sensor nodes is shown in Fig. 59(a), where the estimates are represented by the same color as the respective sensor node.

**Figure 59.** a)The target position estimates by each sensor and b) the target tracks after data fusion

### *9.2.3. Data fusion and target tracking*

The location estimates from all sensors are fused together resulting in a single target location per target. The estimated target locations are not in a track form and contain a significant amount of noise due to small errors in the range estimation. A simplified GM-PHD filter is applied for fusing the target location estimates from all sensors. The location estimates are used to form the observation RFS for the GM-PHD filter. The target state is defined by the 2D target coordinates and velocity vector, **x** = [*x x y*˙ *y*˙] *<sup>T</sup>*. The targets are identified using nearest neighbor association, and the results for the three-target scenario can be seen in Fig. 59(b).

Both methods presented in this section can be used for the localization of multiple non-cooperative targets using distributed UWB sensor network with static sensors as in the scenario described in Section 2.1 .
