Você está na página 1de 9

Passive Infrared Localization with a Probability Hypothesis Density Filter

Jrgen Kemper Robotics Research Institute TU Dortmund University Dortmund University of Technology 44227 Dortmund, Germany Email: juergen.kemper@tu-dortmund.de Daniel Hauschildt Robotics Research Institute TU Dortmund University Otto-Hahn-Str. 8 44227 Dortmund, Germany Email: daniel.hauschildt@tu-dortmund.de

AbstractIn passive infrared localization (PIL) humans are located based on their thermal radiation. Thus, an active tag is not required and privacy is guaranteed due to non-identifying sensors. However, in case of multi-target tracking, the nonidentifying sensors result in missing associations between targets and measurements. As additionally the number of humans in the surveillance region is unknown and varying, conventional localization approaches like the JPDA or MHT lter cannot be applied since they require a xed number of targets or are not tractable. The Probability Hypothesis Density (PHD) lter, on the other hand, that propagates only the rst order moment of the multitarget probability distribution offers an efcient and elegant way to handle the aspects of missing association and varying target number. Furthermore, false measurements as well as missed detection are considered implicitly. In this paper we present an implementation of this lter for PIL that allows an accurate and reliable tracking of several humans. The accuracy and reliability are evaluated under the inuence of noise. Thus, several simulations and real measurements are carried out that reveal a mean error of less than 30 cm in case of three humans. Moreover, efciency tests show that on standard PCs update rates of more than 50 Hz are achievable. Keywords-PHD; Particle Filter; Passive Infrared Localization

I. I NTRODUCTION The ability to locate people is an essential prerequisite to enable location-based services. Therefore, in recent years, several indoor location systems for the eld of ubiquitous computing have been developed. Mostly, these systems apply technologies like active infrared, radio, ultrasound or cameras to gather the required measurements for localization [1][3]. However, none of the common approaches completely meets the special requirements arising in the eld of home automation and Ambient Assisted Living (AAL)1 . On the one hand, systems applying active infrared, radio and ultrasound require carrying an active hardware device, called tag or badge that is able to send or receive some kind of measurement signal. On the other hand, camera based approaches that allow locating people without an additional tag depend
1 http://www.aal169.org

on lighting conditions and result in a violation of privacy. Hence, they suffer from a lack of consumer acceptance. An alternative approach for the eld of home automation and AAL that overcomes these drawbacks was presented by Kemper et al. [3]. It exploits the thermal radiation of humans for localization (passive infrared localization (PIL)). In that case the user is the active source himself. Hence, no tag is required. Furthermore by the usage of low resolution array sensors for detection, privacy is also guaranteed. However, PIL is challenging due to several reasons. First of all, since target identication is not possible, no associations between measurements and humans are given. Furthermore, the number of humans in the surveillance region is unknown and varying. Additionally, due to the limited range of infrared sensors, missed detection can occur. Unfortunately, conventional multi-target tracking (MTT) approaches like MHT2 [4] or the JPDA3 lter [5] that address the data-association problem are either too computationally demanding or cannot be applied as they require a xed number of targets. On the contrary, the Probability Hypothesis Density (PHD) lter invented by Mahler [6] allows a more tractable implementation of MTT since it only propagates the rst order moment of the multi-target posterior. Moreover, it is able to handle a time-varying number of targets, missed detections and false alarms. Although the idea to approximate the multi-target posterior by its rst order moment the PHD is relatively new [6], it has attracted much attention in recent years. First particle lter implementations for applications like radar or terrain tracking, were proposed by Zajic et al. [7], [8], Sidenbladh [9] and Vo et al. [10], [11]. Closed form solutions based on Gaussian mixtures which avoid the problems of state extraction from particles and are much more computationally efcient were invented by Vo et al. [12]. Furthermore, Panta et al. [10], [13] and Clark et al. [14] described enhancements for particle implementations that enable tracking which is not directly addressed by the PHD
2 Multi 3 Joint

Hypothesis Tracking Probability Data Association

lter. Similar enhancements for Gaussian mixture implementations were presented by Clark et al. [15]. In this paper a particle variant (SMC) of the PHD lter is applied to passive infrared localization. Thereby, an efcient positioning and tracking of an unknown and varying number of people is realized. The remainder of this paper is structured as follows. In Sect. II we describe the principles of thermal radiation and infrared sensing. The basics of FISST and the PHD lter are presented in Sect. III followed by the explanation of the lter implementation in case of PIL in Sect. IV. Thereafter, Sect. V describes the results of the evaluation. Finally, Sect. VI draws a conclusion and gives an outlook on future work. II. F UNDAMENTALS OF I NFRARED S ENSING In the following, a short overview on infrared radiation and its sensing is given. A. Radiation Basics Every object with a temperature above absolute zero emits energy via electromagnetic radiation of a broad bandwidth. The overall ux and the wavelength of this radiation depends on its surface temperature T and emissivity . The latter lies in an interval of [0..1] and describes the ability of a surface to emit radiation in comparison to a perfect emitter at the same temperature, a so called blackbody. The emissivity is wavelength dependent, but in most practical cases, where the relevant wavelength range is strongly limited, it can be assumed as constant [16]. The relation between wavelength , temperature T , emissivity and the radiant exitance M is given by Plancks law: M = 5 (eC2 /T () C1 1) (1)

9.44 m. Furthermore, when assuming a relevant temperature range between 0 C and 70 C for indoor localization, the relevant wavelengths of the radiated energy lie in an interval of approximately 8 to 14 m. Thus, the sensitivity of the used sensors may be limited to this range. In this context, a positive side effect is that air has a so called atmospheric window in the interval of approximately 8 to 14 m, as illustrated in Fig. 1. Consequently, the attenuation of the relevant thermal radiation by air is very low over short distances ( 30 m) and thus negligible compared to the attenuation over the distance between object and sensor.
1.0 Transmittance 0.8 0.6 0.4 0.2 0.0 0 1 2 3 4 5 6 7 8 9 Wavelength ( m 10 11 12 13 14 15

O2 H20

H20 CO2 CO2 O3

CO2 O3 Absorbing Molecules

H20

H20 CO2

CO2

Figure 1.

Absorption of thermal radiation in air per seamile [18]

where C1 = 3.74 1012 W/ cm2 and C2 = 1.44 cmK. The total radiated power within a certain range can be calculated by integration, whereas the integral of Formula 1 over cannot be solved analytically in general [17]. An approximation for a broad bandwidth is given by the StefanBoltzmann law: B = A T 4 (2)

Another interesting aspect regarding thermal radiation is that emissivity and absorptivity (characteristics of the surface) are equivalent. For an opaque object the following relationship is valid: Reectivity = 1 . Consequently, a blackbody not only emits at maximum but also absorbs all incident radiation. On the other hand, the real temperature of a highly reective surfaces like polished aluminium ( = 0.04 - 0.06) cannot be detected, but only that of the reected environment. Fortunately, human skin has an emissivity of 0.93 to 0.96 and is thus suitable for thermal detection. B. Detection of Infrared Radiation For the detection of thermal radiation different kinds of sensors exist. They are classied into quantum and thermal detectors. However, the former are not appropriate for PIL as they require cooling (70 K). Additionally, they are strongly wavelength depended. Thermal detectors, on the other hand, need no cooling and have a nearly wavelength independent detectivity. Regarding this class of detectors, the so called thermopiles, offer the best detection capabilities [19], are inexpensive and available as low resolution line and array sensors. Additionally, with response times between 20 50 ms, they are fast enough to track human motion. Thus, they are well suited for appliance in PIL [3]. The working principle of thermopiles is based on a series connection of thermocouples. The resulting junctions are

where = 5.67 108 W/(m2 K4 ), A is the observed surface area and is assumed to be wavelength independent [17]. The wavelength yielding the highest density of radiated power m can be calculated by Wiens law: 2898 m K (3) T with T given in Kelvin. It states that the higher the surface temperature of the radiating object, the shorter is the peak wavelength. Consequently, a human with a skin temperature of 34 C emits most of his radiation near a wavelength of m =

illustrated in Fig. 3 so that multiple maxima can be found. Hence, half as much humans as pixels per sensor can be detected at best so that PIL is limited to track a few users per room. A further disadvantage of this way of detection is that non-human objects that have a similar or higher radiant exitance as the user himself, can tamper the calculated AoA. However, these background effects are not addressed in this paper but were discussed in [3]. III. M ULTI -TARGET T RACKING BASED ON RFS In this section the theoretical foundations of the PHD lter, the Finite Set Statistics (FISST), are explained. Further on, based on these foundations the optimal Bayes recursion for MTT as well as the PHD lter are briey described. A. Finite Set Statistics The Finite Set Statistics offer an elegant way to describe a varying and unknown number of objects. The main idea behind FISST is to convert a localization problem with multiple targets and sensors into a problem with one meta target and one meta sensor [20]. The bundling of sensor and object data is based on a random variable that describes a nite set of mutable length called Random Finite Set (RFS). In case of PIL the RFS of objects consists of a set of realvalued vectors whereas every vector describes the state of one object. Similar to the probability mass function pY (S) = P r(Y S) that describes the statistical behaviour of a random vector Y , the statistical behaviour of a RFS is given by its belief-mass function (S) = P r( S). (4)

Figure 2.

Construction of a thermopile

divided into active and passive ones. The active junctions are attached to a thermally isolated membrane that is exposed to radiation (absorber) as shown in Fig. 2. The passive ones are only inuenced by the ambient temperature. Hence, the output signal of a thermopile depends on the difference between object and ambient radiation (temperature). C. Basics of PIL Localization based on thermopile arrays can be realized by determining the angle of arrival (AoA) under which an object is seen. If several sensors are placed at different locations, the position of a human can lastly be computed via triangulation. If we assume a low resolution line sensor whose pixels exhibit slightly different elds of view, as illustrated in Fig. 3, a simple way of determining the AoA is to choose the direction of that pixel with the highest outcome. Although in case of PIL a more elaborate algorithm is used it is based on the same principle but allows higher resolution and additionally delivers a quality measure, the score. The resulting accuracy equates a standard error of 1.5. The sensor range is limited to approximately 10 m.

Source 2

FoV

Source 1

Figure 3.

Principle of AoA estimation

Moreover, it is also possible to detect several sources, if they are seen by non-adjacent pixels of the sensor as

The overall probability that all elements of an RFS lie in a region S is given by the belief mass [21]. Also on RFSs the mathematical operations of derivation and integration are dened which in that context are called set derivative and set integral. They dene the mathematical connection between belief-mass function and the multiobject probability density f (X): (), (5) f (X)X and f (X) = (S) = X S where X represents a certain instantiation of the RFS . Consequently, based on the set derivative the likelihood functions and markov density distributions can be derived for the case of multiple objects and sensors. The likelihood function f (Z|X), for example, equals the set derivative of the belief-mass function of the corresponding sensor. (|X) (6) f (Z|X) = Z Here, X and Z equate the nite sets of object and measurements, respectively. The interested reader can nd further information on FISST in [22].

B. Multi-target Filter based on RFSs The challenges arising in multi-target localization are often much more complex than in single-target tracking. In case of PIL, for example, the number of objects Mt at time t is not known and can vary due to targets entering or leaving the scene. This is also true for the number of measurements Nt which is uctuating due to false measurements or missed detections. Furthermore, since none identifying sensors are used, measurements cannot be associated with targets. Because of the nite but varying number of targets and objects a description by RFSs is convenient. Consequently (1) (M ) the target states xt , ..., xt t X at time t as well as the (1) (Nt ) measurements zt , ..., zt Z are described by Xt Zt = =
(1) (M ) {xt , ..., xt t } F(X) (1) (N ) {zt , ..., zt t } F(Z),

computational demanding integrations on the state space F(X). C. Probability Hypothesis Density Filter In contrast to the optimal multi-target Bayes lter, the PHD lter only propagates the rst order moment of the posterior, that equates an intensity measure, and is thus less computational demanding. The intensity measure V of an RFS in a region S on the n is dened as [22]: V (S) = E [| S|] =
S

D (x)dx

(11)

and

(7) (8)

Here, D is the density function of also called Probability-Hypothesis-Density (PHD) or intensity. It is given by D = E[ (x)] = whereas X (x) =
wX

X (x) f (X)X,

(12)

where F() describes the collection of all nite subsets of the corresponding subspace. With respect to the changing number of targets pS,t (xt1 ) denes the possibility at time t that xt has already existed at time t 1 (survivors). Consequently, the probability of a vanished target is given by 1 pS,t (xt1 ). Furthermore, pB equals the probability that a new object entered the surveillance region and pG,t (xt1 ) describes the probability of spawning objects. However, in case of PIL spawning can be neglected as humans cannot spawn new targets. Besides, spawning is already handled by the chosen birth model. The RFS Xt of object states lastly results from the union of RFSs of survivors and new born targets. Regarding the measurements, the probability that an object xt is detected is dened by pD,t (xt ). Consequently, the RFS of measurements Zt results from the union of detected objects and false measurements. If the transition density function is given by f (Xt |Xt1 )4 and the likelihood function is represented by g(Zt |Xt ) and if p (Xt |Zt1 ) and p+ (Xt |Zt ) dene the prior and the posterior, respectively, the multi-target Bayes recursion can be expressed as 1. Estimation p = f (Xt |Xt1 ) p+ (Xt1 |Zt1 ) s (dXt1 ), (9) g(Zt |Xt ) p (Xt |Zt1 ) , g(Zt |Xt ) p (Xt |Zt1 ) s (dXt )

w (x),

(13)

with w (x) as the dirac-delta density function concentrated at w. The value of V (S) equals the estimated number of elements whose positions are given by the peaks of D . Under the assumption, that the RFSs of false measurements and objects are Poisson, the PHD lter can be described by the following recursion: 1. Estimation D (xt |Zt1 ) = + + pS,t (xt1 ) f (xt |xt1 ) D+ (xt1 |Zt1 )dxt1 (xt |xt1 ) D+ (xt1 |Zt1 )dxt1 (xt ) (14)

The rst term species the share of the survivors, the second one the share of spawned and the third one the share of the new born objects, whereas and represent the corresponding intensity functions. As already stated, in case of PIL spawning can be neglected 2. Correction D+ (xt |Zt ) = [1 pD,t (xt )] D (xt |Zt1 ) +
zt Zt

2. Correction p+ = (10)

where s describes an appropriate measure on F(X) (see [11]). Although by now several SMC implementations of this recursion were presented [11], [23], the multi-target Bayes lter it is not tractable in general, as it requires several
4 At this point markov states are assumed. However, this is not a necessary assumption.

pD,t (xt ) gt (zt |xt ) D (xt |Zt1 ) (15) (zt ) + K pD,t (xt ) gt (zt |xt ) D (xt |Zt1 )dxt

with K =

Here, the rst term species the share of missed detections and the second one that of detected objects. Moreover, the

normalizing denominator results from the sum of the intensity function of false measurements and the integration of the nominator over the surveillance region. However, this correction equation is only valid for one sensor. In case of several sensors the calculation is typically done sequentially for every sensor which is described later in more detail. IV. PASSIVE I NFRARED L OCALIZATION WITH A SMC-PHD F ILTER In the following a particle implementation of the PHD lter for PIL is described. But rst of all the used motion and sensor models are explained. A. Motion Model When localizing humans no control data is available that gives direct information about the state transition. Thus, the only way to generate corresponding information is to estimate the target velocity based on the former movement. The resulting motion model is given by Eqn. 16 xt = xt1 + vt + N(0, a2 t2 ) t, (16) where vt = (xt1 xt2 )/t with an additional zeromean and Gaussian white noise N(). Furthermore, as higher speeds are not very likely indoors, the motion is constrained in two ways. The maximum velocity is limited to 3 m/s and the maximum acceleration is set to 1 m/s2 Regarding the state transition, a further aspect that has to be modelled is the generation of new objects. Beside the possibility to dene certain areas close to a door where new targets can enter the surveillance region, an additional option is to realize the particle generation with respect to the current measurements. Here, the latter is chosen due to the applied sensor conguration. Hence, around all intersections of two of the measured AoAs a xed number of new particles is generated randomly according to a Gaussian distribution. B. Sensor Model As already explained, the measured angles exhibit a standard error of approximately 1.5. Consequently, the sensor is modelled by the given measurement values and an additional zero-mean white Gaussian noise. y mS,y z = arctan (17) S + N(0, 2 ), x mS,y where the target location is represented by (x,y) and the sensor position and orientation by mS and S , respectively. Moreover, as the sensors are noisy, false measurements can occur. However, these measurements can be rejected based on the corresponding score. That means that measurements with a score under a certain threshold are not considered. Thus, false measurements can be neglected in this realization. Another aspect that has to be considered is the limited sensor range which leads to missed detections. Thus, the

detection probability depends on the room size, but also on the target positions, as users seen by adjacent pixels cannot be discriminated. Therefore, a sensor specic detec[r] tion probability pD,t is dened as the product of the room dependent probability pgen and the measurement specic D,t detection probability psen . In order to determine the latter, D,t it is necessary to estimate the number of objects. This estimation is realized in the following way: If we assume that the FoV of every sensor or a combination of several sensors at the same position covers the whole room, the number of objects can be approximately calculated by MtD = pgen D,t NS
NS

r=1

N MS,r ,

(18)

where N S and N MS,r equal the number of sensors and measurements of the current sensor, respectively. Thereby, the sensor specic detection probability equals pD,t =
[r]

N MS,r MtD

with r = 1, ..., N S

(19)

which relates the current number of measurements with the estimated number of objects. C. SMC-PHD-Filter Based on the described motion and sensor models in this section a multi-sensor SMC implementation of the PHD lter is presented that allows the approximation of arbitrary distributions by a set of weighted particles whereas every particle represents one state hypothesis. The weight of a particle corresponds to the probability of this hypothesis. In case of the PHD lter the weights of all particles sum approximately to the number of estimated targets. However, in addition to the recursion given by Eqn. 14 and 15 that only consists of estimation and correction two further steps, the initialization and the resampling, are required. Initialization: During initialization which is only executed once in the beginning a certain number of particles that serve as starting point for the lter recursion is distributed due to an initial distribution, e. g. randomly, in the surveillance region. The accumulated weights of these particles correspond to the initially assumed number of targets. Estimation: The estimation consists of two steps, the motion update of the N old particles including the reweighting with the survival probability and the generation of L new particles according to the birth model. Assuming that the latter are represented by the distribution qt (|Zt ), the new particle set t , approximating the prior D (xt |Zt1 ) is given by
t

xt , w t

(i)

(i)

| i = 1, ..., N + L

where the particle positions are represented by xt


(i)

D. State extraction In order to determine the object positions from the resulting particle set, Mt accumulation points (clusters) with the highest number of particles have to be identied. Applicable techniques for this extraction are the Expectation Maximum (EM) algorithm or the K-means clustering. Clark and Bell showed that in the context of PHD ltering Kmeans clustering outperforms EM [14]. The idea of the K-means is to assign all particles to the nearest of a given number of cluster centers. The new centers are then calculated by averaging over all particle positions of one cluster. This process is repeated iteratively until the variance of cluster centers falls under a certain threshold. In order to minimize the required iterations, the initial centers should be near the real ones. Thus here, the initial centers are chosen with respect to the old positions as described in the next subsection. E. Tracking In many application not only the positions but also the walked track is of some relevance. However that requires an assignment of new positions to existing tracks. e. g. by determining the nearest neighbour, as tracking is not supported directly by the PHD lter. In case of this realization tracking is of signicant importance as the current target velocity is estimated based on the last track positions (see Sect. IV-A). Therefore, the particles are given an additional label that assigns them to a certain track so that during motion update the corresponding target velocity can be considered. Several approaches to realize target tracking for PHD ltering were presented by Panta et al. [13] and Clark et al. [14] like particle labelling, estimate-to-track association or a mixture of MHT and PHD-lter. In this realization particle labelling is used due to the motion update and the track association is realized by a nearest-neighbour search. Thus, the applied procedure establishes a combination of Particle Labeling and Estimate-to-Track association as described by Clark et al [14]. The resulting algorithm is briey described in the following: In the beginning the initial cluster centers for the current iteration are calculated. If the number of objects did not change since the last iteration, the centers equate the old centers propagated over time with the corresponding track velocities. In case of a changed number of objects Mt initial centers are ascertained by a grid based partitioning of the room, whereas every particle is associated to one grid element. Then the Mt elements with the highest number of particles are chosen and ordered with respect to the old but propagated object states by a nearest-neighbour assignment. Dependent on the difference between new and old number of objects, tracks are added or deleted if necessary. Afterwards, k-means clustering is processed and the resulting positions are used to update the tracks, whereas the association is given implicitly due to the former processed ordering.

f (|xt1 ) q(|Zt )
(i) (i)

(i)

with i = 1, ..., N with i = N + 1, ..., N + L with i = 1, ..., N with i = N + 1, ..., N + L

and the corresponding weights are given by wt =


(i)

pS,t (xt1 ) wt1


pB,t L

Here, L equates the product of particles per and the number of intersections. Moreover, it is assumed that only one object can be born per time. Correction: During the correction step, all weights of t are recalculated based on an adapted version of Eqn. 15 that considers more than one sensor. Hence, the following calculation is executed iteratively over all R sensors, whereas the PHD resulting from the former sensor replaces D (xt |Zt1 ), except for the rst sensor5 . In case of this (i) implementation, it means that wt equals the particle weight of the last iteration [24]: wt
(i)

= +

(1 pD,t ) wt
[r] [r]

[r]

(i) (i) (i)

pD,t g [r] (zt |xt ) wt


[r] zt Zt

t (zt ) + pD,t Gt (zt )

[r]

[r]

(20)

with G (zt ) =
i=1 [r] N +L

g [r] (zt |xt ) wt

(i)

(i)

(21)

for all r = 1, ..., R. Additionally, the equation considers that the sensor model and the detection probability can be different for every sensor. Resampling: The subsequent resampling creates a new + particle set t by sampling particles from the current one, whereas the current particle weights dene the probability that a particle is chosen. Hence, a particle with a high weight can be sampled several times as well as particles with low probability may be completely ignored. Lastly, the resulting set consists of equally weighted particles that agglomerate at locations of high intensity. As furthermore the weight sum of the new set should correspond to the estimated number of objects Mt , before resampling this number has to be computed based on t according to Eqn. 22
N +L

Mt =
i=1

wt ,

(i)

(22)

with subsequent rounding to the nearest integer. If the number of objects has changed compared to the last iteration, also the amount of particles has to be adjusted, as a xed number of particles is used per object. Thus, the particle Mt weights are given by N + , where N + equates the new number of particles.
5 In this iteration the order of processing inuences the result, though it can be neglected in practice [22]

100 90 80 Position error [cm] 70 60 50 40 30 20 10 0 0 1 Sensor 2 3 Square Circle Position error [cm]

100 90 80 70 60 50 40 30 20 10 0 0 1 Sensor 2 3 Square Circle

(a) Walking in same direction Figure 4.

(b) Walking in opposite direction Accuracy for two humans with respect to sensor noise

V. E VALUATION In the following, the developed lter is evaluated under the aspects of accuracy and efciency. Thus several simulation results and some real measurements are presented. In both cases a detection probability of 0.75 is assumed and every object is approximated by 500 particles. Furthermore, at every potential birth position 30 particles are generated whereas the survival and birth probability are set to 0.95 and 0.05, respectively. A. Simulations The simulated testbed consists of four sensor modules that are placed at chest height in the corners of a room. Every module exhibits two thermopile line sensors that in combination enable a eld of view (FoV) of 90. For the sensor simulations a proprietary OpenGL simulator is used, whereas the sensor parameters are chosen in a way that the outcomes are comparable to real measurements. The results for two humans walking in the same/opposite direction (see also Fig. 6(a)) one on a square and one on a circle trajectory are illustrated in gure 4 for different noise levels6 . Accordingly, the absolute position error is approximately 20 cm and nearly independent of noise. Moreover, the maximum position error is less than 40 cm for similar walking directions and less than 1 m in the opposite case. The higher maximum error in the latter simulation is caused by discrimination problems when the humans meet, as Fig. 6(a) illustrates. At this moment only a combination of both humans is modelled for a certain time. Further simulations for one and three humans yield position accuracies of 15 cm and 25 cm, respectively. Additional investigations were made to test the estimation accuracy of the number of targets. In case of one and two
6 The noise is given with respect to the resolution of the AD converter of the sensor module, whereas the typical sensor noise is approximately 1.75 (8.5 mV).

targets this estimation was always correct. When simulating three targets on the other hand, the number was sometimes overestimated as illustrated in Fig. 5. Lastly, it should be clear that estimation errors cannot be avoided and become more likely with an increasing number of objects. However, in intended applications, typically not more than three people must be located per room.
5 4 no. of humans 3 2 1 0 0 20 40 60 80 100 time [s]

Figure 5.

Estimation error for three humans

B. Measurement Results Real measurements were carried out to conrm the simulation results. Thus, a testbed of 4.9 6.2 m with for double sensors in the room corners was constructed, as depicted in Fig. 6(b). Again, tests with different trajectories (square, diagonal) were processed. The results are presented in Tab. I. However, these errors do not present the absolute position errors but the minimum distance to the walked track as a reference system was not available. The real position errors are higher. This can also be seen in Fig. 6 that opposes the calculated tracks of simulation and real measurements for two targets and illustrates that the real measurements exhibit a higher error due to several reasons.

S1

S4

S1

S4

4,9 m

2,9 m

5m

concourse

S2 5m
(a) Simulations (opposite direction) Figure 6.

S3

S2

4,2 m 6,2 m
(b) Real measurements

S3

Resulting track for localization of two humans

First of all, the simulation does not consider component tolerance of sensors, especially of the optics. That is, sensors with equal optical and electrical characteristics are assumed. Moreover, in contrast to the simulation of a human does not always walk on exactly the same trajectory. However lastly, it can be stated that the measurements conrm the simulation results.
minimum error One person Two persons 9 cm 12 cm max. error 26 cm 68 cm

Table I T RACK LOCALIZATION ACCURACY FOR ONE AND TWO HUMANS

C. Runtime Finally, the runtime of the approach is briey evaluated. In general it can be stated that due to the sequential sensor processing the runtime is dominated by the correction step and additionally by the clustering. Thus, it depends on the number of objects M , particles N , sensors S and clustering iterations I: O(N (S + M I)) Exemplary tests on a standard dual-core PC with 3 GHz yield an update rate of 50 Hz for three objects and 1500 particles which is more than fast enough for PIL. Moreover, it is likely that further improvements can be achieved by code optimization. VI. C ONCLUSION In this paper we presented an SMC implementation of the PHD lter for passive infrared localization that allows the simultaneous positioning and tracking of several people based on their thermal radiation. The benet of the used PHD lter is that the number of targets which in case of PIL

is unknown and varying is implicitly estimated. Furthermore, aspects like false measurements and missed detection are considered. Simulations show that typically up to three humans can be discriminated with a mean position error of less than 25 cm. Real measurements conrm the results but are slightly worse as simulations do not consider component tolerances of the used sensors. Furthermore, runtime performance tests show that on a standard PC update rates of 50 Hz can be realized which is sufcient for PIL. Future work will be concentrated on a more efcient implementation of the lter based on Gaussian mixtures as presented by Vo et al. [12]. Furthermore, as already described by Kemper et al. [3] background radiation has to be identied and ltered in order to enable the appliance in typical indoor environments that may exhibit different sources of disturbance. R EFERENCES
[1] J. Hightower and G. Borriello, Location Systems for Ubiquitous Computing, Computer, vol. 34, no. 8, pp. 5766, August 2001. [2] J. A. Tauber, Indoor Location Systems for Pervasive Computing, Massachusetts Institute of Technology, Cambridge, MA, USA, Tech. Rep., 2002. [3] J. Kemper and H. Linde, Challenges of Passive Infrared Indoor Localization, in 5th Workshop on Positioning, Navigation and Communication (WPNC), March 2008. [4] S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics, 1st ed. MIT Press, 2005. [5] J. Vermaak, S. J. Godsill, and P. Perez, Monte Carlo Filtering for Multi-target Tracking and Data Association, IEEE Transactions on Aerospace and Electronics Systems, vol. 41 (1), pp. 309332, January 2005.

[6] R. P. S. Mahler, Multitarget Bayes ltering via rst-order multitarget moments, vol. 39, no. 4, pp. 11521178, Oct. 2003. [7] R. P. S. Mahler and T. Zajic, A particle-systems implementation of the phd multitarget tracking lter, in Processing, Sensor Fusion and Target Recognition, Proceedings of SPIE, I. Kadar, Ed., vol. 5099, no. XII, 2003, pp. 291293. [8] R. P. S. M. T. Zajic, Multi-Object Tracking Using a Generalized Multi-Object First-Order Moment Filter, in Proc. Conference on Computer Vision and Pattern Recognition Workshop CVPRW 03, vol. 9, 1622 June 2003, pp. 9999. [9] H. Sidenbladh, Multi-Target Particle Filtering for the Probability Hypothesis Density, in 6th International Conference on Information Fusion, Cairns, Australia, 2003, pp. 800806. [10] K. Panta, B. Vo, and S. Singh, Improved probability hypothesis density (phd) lter for multitarget tracking, in Third International Conference on Intelligent Sensing and Information Processing, Dec. 2005, pp. 213218. [11] B.-N. Vo, S. S. Singh, and A. Doucet, Sequential Monte Carlo methods for Multi-target Filtering with Random Finite Sets, IEEE Transactions on Aerospace and Electronic Systems, vol. 41, no. 4, pp. 12231245, October 2005. [12] B.-N. Vo and W.-K. Ma, The Gaussian Mixture Probability Hypothesis Density Filter, Signal Processing, IEEE Transactions on, vol. 54, no. 11, pp. 40914104, November 2006. [13] K. Panta, B.-N. Vo, and S. Singh, Novel data association schemes for the probability hypothesis density lter, Aerospace and Electronic Systems, IEEE Transactions on, vol. 43, no. 2, pp. 556570, April 2007. [14] D. E. Clark and J. Bell, Multi-target state estimation and track continuity for the particle PHD lter, Aerospace and Electronic Systems, IEEE Transactions on, vol. 43, no. 4, pp. 14411453, October 2007.

[15] D. E. Clark, K. Panta, and B.-N. Vo, The Gaussian mixture PHD lter Multiple Target Tracker, in In Proceedings 9th Annual Conference Information Fusion, 2006. [16] Everett Infrared, Physics of Electro-Optic Detectors, 2005. [Online]. Available: www.everettinfrared.com/detectors.pdf [17] J. Fraden, Handbook of Modern Sensors - Physics, Designs and Applications, 3rd ed. Springer-Verlag, 2004. [18] H. A. Gebbie, W. R. Harding, C. Hilsum, A. W. Pryce, and V. Roberts, Atmospheric Transmission in the 1 to 14 Region, Royal Society of London Proceedings Series A, vol. 206, pp. 87107, Mar. 1951. [19] HAMAMATSU PHOTONICS K.K.- Solid State Division, Technical Information SD-12 Characteristics and use of infrared detectors, November 2004. [20] R. P. S. Mahler, Statistics 101 for Multisensor, Multitarget Data Fusion, IEEE A&E Systems Magazine, vol. 19, no. 1, pp. 5364, January 2004. [21] R. R. S. Mahler, Engineering statistics for multi-object tracking, in Proc. IEEE Workshop on Multi-Object Tracking, 8 July 2001, pp. 5360. [22] R. P. S. Mahler, Statistical Multisource-Multitarget Information Fusion. Artech House, Boston, 2007. [23] H. Sidenbladh and S.-L. Wirkander, Tracking Random Sets of Vehicles in Terrain, in Proc. Conference on Computer Vision and Pattern Recognition Workshop CVPRW 03, vol. 9, 2003, pp. 9898. [24] M. Tobias and A. D. Lanterman, Probability hypothesis density-based multitarget tracking with bistatic range and Doppler observations, IEEE Proceedings -Radar, Sonar and Navigation, vol. 152, no. 3, pp. 195205, 2005.

Você também pode gostar