Beruflich Dokumente
Kultur Dokumente
September 2003
1
D E A L I N G W I T H U N C E R TA I N T Y
L
ocation awareness is important to Bayes filters
many pervasive computing applica- Bayes filters2 probabilistically estimate a
tions. Unfortunately, no location sen- dynamic system’s state from noisy observations.
sor takes perfect measurements or In location estimation for pervasive computing,
works well in all situations. Thus, the the state is a person’s or object’s location, and
motivation behind this article is twofold. First, we location sensors provide observations about the
believe the pervasive computing community will state. The state could be a simple 2D position or
benefit from a concise survey of Bayesian-filter tech- a complex vector including 3D position, pitch,
niques. Because no sensor is perfect, representing roll, yaw, and linear and rotational velocities.
and operating on uncertainty with a statistical tool Bayes filters represent the state at time t by ran-
such as Bayes filters is key in any dom variables xt. At each point in time, a prob-
Dieter Fox, Jeffrey Hightower, system using many sensors. ability distribution over xt, called belief, Bel(xt),
Lin Liao, and Dirk Schulz
Second, estimating an object’s represents the uncertainty. Bayes filters aim to
University of Washington
location is arguably the most sequentially estimate such beliefs over the state
Gaetano Borriello fundamental sensing task in space conditioned on all information contained in
University of Washington and many pervasive computing sce- the sensor data.
Intel Research Seattle narios. It is thus a natural do- To illustrate, let’s assume that the sensor data
main in which to illustrate the consists of a sequence of time-indexed sensor
application of Bayesian filter observations z1, z2, …, zt. The belief Bel(xt) is then
techniques. Representing locations statistically defined by the posterior density over the random
enables a unified interface for location informa- variable xt conditioned on all sensor data avail-
tion. This lets us write applications independent able at time t:
of the sensors used—even when using very dif-
ferent sensor types, such as GPS and infrared Bel(xt)= p(xt | z1, z2, …, zt). (1)
badges. (A comparative survey of location sys-
tems appears elsewhere.1) Roughly speaking, the belief answers the ques-
Here, we illustrate fusing sensor data from ultra- tion, “What is the probability that the person is at
sound and infrared tags. We also discuss how to location x if the history of sensor measurements
combine high-resolution location information is z1, z2, …, zt?” for all possible locations x. In gen-
from anonymous laser range finders with low-res- eral, the complexity of computing such posterior
olution location sensors that provide identification. densities grows exponentially over time, because
24 PERVASIVE computing Published by the IEEE CS and IEEE ComSoc ■ 1536-1268/03/$17.00 © 2003 IEEE
Bel – (x 0 )
(a) x
p(z | x)
x
Bel (x 0 )
(b) x
Bel – (x 1 )
x
(c)
p(z | x)
x
Bel (x 1 )
x
(d)
Bel – (x 2 )
x
(e)
Figure 1. A one-dimensional illustration of Bayes filters. A person carries a door-sensing camera that cannot distinguish different doors.
Each frame depicts the person’s position in the hallway and the current belief Bel(xt): (a) The person’s location is unknown. (b) The
sensor sends a “door found” signal. (c) The person moves. (d) The sensor observes another door. (e) The person moves again.
Additionally, (b) and (d) depict the observation model p(z|x), the probability of observing a door at different locations in the hallway.
the number of sensor measurements an object’s location at time t depends only hypothetical one-dimensional scenario,
increases over time. To make the com- on the previous state xt–1. States before a person walks down a hallway carrying
putation tractable, Bayes filters assume x t–1 provide no additional information. a sensor (such as a camera) that signals
the dynamic system is Markov—that is, Under the Markov assumption, we if the person walks by a door. The sensor
the current state variable xt contains all can efficiently compute the belief in cannot distinguish different doors. We
relevant information. For locating objects, Equation 1 without losing information. are not suggesting this mobile camera
the Markov assumption implies that sen- Before we provide the update equations, scenario is a viable location system
sor measurements depend only on an let’s briefly examine the recursive Bayes implementation; it simply illustrates
object’s current physical location and that filter update steps using Figure 1. In this important Bayes filters properties.
As Figure 1a illustrates by showing a Here, p(xt | xt–1) describes the system context of location estimation for per-
uniform distribution over possible loca- dynamics—that is, how the system’s vasive computing.
tions, the person’s location is at first state changes over time.
unknown. The sensor then sends a In location estimation, this conditional Kalman filters
“door found” signal. The resulting belief probability is the motion model—where Kalman filters are the most widely used
places high probability at places next to the object might be at time t, given that variant of Bayes filters.3 They approxi-
doors and low probability elsewhere (see it was previously at location xt–1. This mate beliefs by their first and second
Figure 1b). This distribution possesses model strongly depends on the informa- moment, which is virtually identical to a
three peaks, each corresponding to one tion available to the estimation process. unimodal Gaussian representation:
of the environment’s (indistinguishable) It can range from predicting the next
doors. Furthermore, the resulting distri- position by estimating a person’s motion Bel( xt ) ≈ N ( xt ; µ t , Σ t )
bution assigns high probability to three velocity to predicting when a person will 1
=
distinct locations, illustrating that the exit the elevator by estimating the per- d /2
(2π ) Σ t
1/ 2
probabilistic framework can handle mul- son’s goal.
1
tiple, conflicting hypotheses that natu- In the hallway example, the motion exp − ( xt − µ t )T Σ t−1( xt − µ t ).
rally arise in ambiguous situations. update corresponds to the change in 2
Finally, even nondoor locations possess belief leading to Figure 1c and 1e. The (4)
nonzero probability because of the un- belief immediately after the prediction
certainty inherent in sensing; with a and before the observation is the predic- Here, µt is the distribution’s mean (first
small but nonzero probability, the sen- tive belief Bel−(xt). The perceptual model, moment) and Σt is the d × d covariance
sor might err and actually not be next to p(zt | xt), describes the likelihood of mak- matrix (second moment), where d is the
a door. ing observation zt given that the person is state’s dimension. N(xt; µt, Σt) denotes
Figure 1c shows the person moving at location xt. As Figure 1b and 1d show, the probability of xt given a Gaussian
and this motion’s effect on the belief, the observation update rule increases the with mean µt and covariance Σt, as Equa-
assuming the person moves to the right probability of locations with high obser- tion 4 specifies. Σt represents the uncer-
with typical walking speed. The Bayes vation likelihoods. For location estima- tainty in the estimate—the larger the
filter shifts the belief in the direction of tion, the perceptual model is usually con- covariance, the wider the distribution’s
motion and smoothes it out to account sidered a given sensor technology’s spread. Kalman filters are optimal esti-
for inherent uncertainty in motion esti- property. It depends on the sensors’ types mators, assuming the initial uncertainty
mates. Finally, Figures 1d and 1e depict and positions and captures their error is Gaussian and the observation model
the belief after the sensor observes characteristics. In Equation 3, αt is sim- and system dynamics are linear functions
another door and after the next motion, ply a normalizing constant that ensures of the state. Because most systems are
respectively. Most of the probability that the posterior over the entire state not strictly linear, people typically apply
mass is placed on a location near one of space sums up to one. Bel(x0) is initial- extended Kalman filters, which linearize
the doors, and the filter is now quite con- ized with prior knowledge about the the system using first-order Taylor series
fident of the person’s location. object’s location, typically uniformly dis- expansions.3
Here’s how to update the Bayes filter. tributed if no prior knowledge exists. Kalman filters’ main advantage is their
Whenever a sensor provides a new Bayes filters are an abstract concept in computational efficiency. We can imple-
observation zt, the filter predicts the state that they provide only a probabilistic ment both the prediction (Equation 2)
according to framework for recursive state estimation. and correction (Equation 3) using effi-
Implementing Bayes filters requires spec- cient matrix operations on the mean and
Bel − ( xt ) ← ifying the perceptual model p(zt | xt), the covariances. This efficiency, however,
dynamics p(xt | xt–1), and the representa-
∫ p( xt xt −1)Bel( xt −1)dxt −1. (2)
tion of the belief Bel(xt). The properties
comes at the cost of restricted represen-
tational power because Kalman filters
of the different implementations of Bayes can represent only unimodal distribu-
The filter then corrects the predicted esti- filters strongly differ in how they repre- tions. So, Kalman filters are best if the
mate using this sensor observation: sent probability densities over the state xt. uncertainty in the state is not too high,
We now discuss different belief repre- which limits them to location tracking
Bel( xt ) ← α t p( zt xt ) Bel − ( xt ). (3) sentations and their properties in the using either accurate sensors or sensors
The MHT tracks each Gaussian update it for every new observation.
hypothesis using a Kalman filter. The
technique determines the hypotheses’ tributions over the discrete state space spaces. Their disadvantage is the repre-
weights wt( i) on the basis of how well the and can solve estimation problems such sentation’s coarseness. Estimates provide
hypotheses predict the sensor measure- as the one in Figure 1. The mobile-robot- only rough information about a person’s
ments. In other words, at each update, localization community has shown that location. Topological approaches are
the weights are set proportional to the metric approximations provide accurate typically adequate if the sensors in the
likelihoods of the sensor measurements, position estimates in combination with environment provide only very impre-
given the individual hypotheses. high robustness to sensor noise.5 Grid- cise location information.
Owing to MHT approaches’ ability to based approaches’ disadvantage is the
represent multimodal beliefs, they are computational and space complexity Particle filters
more widely applicable than the Kalman required to keep the position grid in Particle filters represent beliefs by sets
filter. For example, in contrast to Kalman memory and to update it for every new of samples, or particles:7
filters, we can apply MHT approaches observation. Because the complexity
to the problem illustrated in Figure 1.
However, MHT techniques are compu-
grows exponentially with the number of
dimensions, we can apply grid-based
Bel ( x t ) ≈ S t = {x (i ) (i )
t ,w t }
i = 1, ..., n .
(6)
tationally more expensive and require approaches only to low-dimensional
sophisticated techniques or heuristics to estimation problems, such as estimating Here, each xt( i) is a state, and the wt( i) are
determine when to add or delete hypo- a person’s position and orientation. nonnegative weights called importance
theses. Because each hypothesis is tracked factors, which sum up to one. Particle
using a Kalman filter, these methods still Topological approaches filters realize Bayes filter updates accord-
rely on the linearity assumptions of We can avoid the computational com- ing to a sampling procedure, often called
Kalman filters. In practice, however, mul- plexity of grid-based methods using non- sequential importance sampling with
tihypothesis approaches have been metric representations of an environment. resampling. (An overview of particle fil-
robust to violations of these assumptions. For instance, many indoor environments ters and various applications appears
provide a natural way to represent a per- elsewhere.7)
Grid-based approaches son’s location at a symbolic level such as Figure 2 illustrates the particle filter
These approaches overcome the restric- the room or hallway the person is cur- algorithm using our one-dimensional
w (x )
(a) x
p(z | x )
x
w (x )
x
(b)
w (x )
x
(c)
p(z | x )
x
w (x )
x
(d)
w (x )
x
(e)
Figure 2. Applying particle filters to location estimation. The black bars depict the particles representing the belief Bel(xt). (a) A
uniformly distributed sample set presents the person’s initially unknown position. (b) A sensor detecting the left door. The sample
set is obtained from weighing the importance factors in proportion to the likelihood of the measurement. (c) An implementation
of the prediction step. The samples were drawn from the previous set with probability proportional to the importance factors.
(d) A sensor detecting a second door. (e) The sample set obtained after another prediction step.
hallway example. A uniformly distrib- surement by adjusting and normalizing When the person moves to the right,
uted sample set represents the person’s each sample’s importance factor, lead- particle filters randomly draw samples
initially unknown position (see Figure ing to the sample set in the lower part from the current sample set (with prob-
2a). Each sample has the same impor- of Figure 2b. These samples are the ability given by the importance factors).
tance factor w(x), as the equal heights same as before, but now their impor- Then the filters use the model to guess
of all bars in Figure 2a indicate. In Fig- tance factors are proportional to the the possible succeeding location for each
ure 2b, the sensor detects the left door. observation likelihood p(z | x), shown new particle. This update implements
The particle filter incorporates the mea- in Figure 2b. the general Bayes filter’s prediction step
Comparing Bayes filters implementations (+, 0, and – represent good, neutral, and weak, respectively).
Kalman Multihypothesis Grid Topology Particle
tracking
(Equation 2). Figure 2c shows the result- how well the different approaches can MIT’s Cricket ultrasound tags, and the
ing sample set, which differs from the estimate location given adequate sensors. LMS200 180-degree scanning laser
original one in that most samples center Obviously, grid-based approaches can range finder from SICK, Inc. (Numerous
around three locations. This concentra- reach arbitrary accuracy but at prohibi- references to the seminal work in sensors
tion of the samples is achieved through tively high computational costs. Kalman for location appears elsewhere.1) The
sampling proportional to the weights. filters’ limited robustness is due to the sensors are deployed throughout the
In Figure 2d, the sensor detects a sec- unimodal belief representation. Intel Research laboratory in Seattle.
ond door, leading to the likelihood p(z | Both Kalman filters and MHT require
x). By weighing the importance factors in accurate sensors with rather high update Sensor models
proportion to this probability, we obtain rates. Topological approaches require sen- To apply Bayes filters to location esti-
the sample set shown in the lower part of sors that relate to an environment’s lay- mation using a specific sensor type, we
Figure 2d. After the next prediction out. Grid-based approaches and particle must first generate a sensor model p(z |
update, most of the probability mass is filters can incorporate virtually any sensor x) describing the likelihood of observing
consistent with the person’s true loca- type. Kalman filters are the most efficient a sensor measurement z given the person
tion. So, the result is consistent with the in terms of memory and computation. or object’s state x. Such a model consists
general Bayes filter example in Figure 1. Efficiency is a key disadvantage of of two types of information: a map of the
Particle filters’ key advantage is their grid-based methods, although tree-based environment and the sensor noise. The
ability to represent arbitrary probability implementations can reduce this prob- problem of constructing maps of indoor
densities. Furthermore, unlike Kalman lem.5 In general, if accurate sensors are environments has received substantial
filters, particle filters can converge to the available, Kalman filters might be the attention in the robotics research com-
true posterior even in non-Gaussian, non- best choice. For specific sensors, and if munity.9 Figure 3 illustrates sensor mod-
linear dynamic systems. Compared to accurate location estimates are not els for ultrasound tags, infrared badges,
grid-based approaches, particle filters are required, topological approaches pro- and laser range finders, respectively.
very efficient because they automatically vide a good way to estimate a person’s Figure 3a shows the likelihood of
focus their resources (particles) on location. Otherwise, particle filters are observing a 4.5-meter ultrasound mea-
regions in state space with high proba- an extremely flexible tool with low surement for the different locations in
bility. Because particle filters’ efficiency implementation overhead. the environment. Because such time-of-
strongly depends on the number of sam- flight sensors provide information about
ples used for filtering, several improve- Example applications the distance between the sensor and the
ments have been made to more efficiently Here, we show sensor fusion of two person, the likelihood function is a ring
use the available samples.8 However, representative pervasive location tech- around the sensor’s location. The ring’s
because these methods’ worst-case com- nologies: infrared proximity badges and width represents the uncertainty in the
plexity grows exponentially in the dimen- ultrasonic time-of-flight tags. Then we measured distance and is often repre-
sions of the state space, we must be care- present a novel approach to tracking sented by a Gaussian distribution cen-
ful when applying particle filters to multiple objects that combines the accu- tered at the measured distance. Fur-
high-dimensional estimation problems. racy benefits of anonymous sensors such thermore, because ultrasound sensors
as scanning infrared laser range finders frequently produce measurements that
Comparison with the identification certainty of the are far from the true distance owing to
Table 1 summarizes the advantages less accurate infrared and ultrasonic ID reflections, all locations in the free space
and disadvantages of different Bayes fil- sensors. Our hardware is the commer- have a nonzero likelihood, as the map’s
ter implementations. Accuracy measures cial VersusTech infrared badge system, gray areas indicate.
Figure 3. Probabilistic models for (a) ultrasound tags, (b) infrared badges, and (c) laser range finders. The environment is 30 m × 30 m.
In (a) and (b), darker areas indicate higher likelihoods of observing the corresponding measurement. Walls and other obstacles are
white because they have zero likelihood. In (c), laser-scan beams, along with a detected person, are blue. The person obstructs the laser
beams, thereby causing a “shadow” in the scan.
Figure 3b illustrates the sensor model ity. Bayes filters can naturally integrate efficiently by constraining the possible
for the infrared badge system. Such sen- information from different sensors. In locations. More specifically, we could con-
sors provide no distance information— this case, whenever an ultrasound or strain the state space to locations on a
only information about a person’s pres- infrared sensor detects the person, the Voronoi graph, which is a structure simi-
ence within a certain range of the particles are weighted with observation lar to a skeleton of an environment’s free
receiver. Accordingly, an infrared mea- likelihoods such as in Figures 3a and 3b, space.10 The key advantage of such graphs
surement has a circular likelihood respectively. is that they naturally represent typical
around the sensor location. Figure 4 shows snapshots from a typ- motion along the environment’s main
Figure 3c shows a scan taken from a ical sequence projected onto a map of the axes. This technique differs from a topo-
laser range finder. Because these sensors environment. The person is wearing an logical approach, however, because it
are at fixed positions, detecting sensor infrared badge and ultrasound tag and maintains high precision in the location
beams that return atypically short mea- starts in the upper-right corner as the icon estimate—the particles can flow freely
surements is straightforward. Several indicates (see Figure 4a). Because the sys- along the graph’s edges in contrast to the
adjacent short beams form a shadow tem doesn’t know the start location, the discrete probabilities stored only at the
region indicating a person’s presence, as samples are spread uniformly through- graph nodes in a topological approach.6
Figure 3c shows. A Gaussian distribution out the environment’s free space. Figure Figure 5 shows a sample run using par-
usually represents the likelihood for such 4b shows the belief after the person ticle filters on Voronoi graphs. The lines
detected features with small uncertainty. moves out of the cubicles, at which point in the first frame indicate the Voronoi
the samples are spread over different graph. The sequence is based on data
Multiple inaccurate ID sensors locations owing to the inaccurate sensor identical to that in Figure 4 using uncon-
Here we illustrate how to estimate a information received so far. After an strained particle filters. In experiments we
person’s location using multiple inaccu- ultrasound sensor detects the person, the found that such Voronoi graph tracking
rate ID sensors, such as the ultrasound filter can estimate the location more accu- produces better estimates using far fewer
and infrared badge systems in our test rately (see Figure 4c). The estimates’ cer- particles. Furthermore, we can use the
environment. Owing to these sensors’ tainty continues to change depending on Voronoi graph structure to learn a per-
high noise level, the belief over the per- the sensor measurements’ quality. son’s high-level motion patterns, such as
son’s location is typically uncertain and This example shows that particle filters “person A typically goes into the printer
multimodal; so, we chose particle filters can extract reasonable location informa- room when she walks down hallway 3.”
for this application. The state vector x tion even from inaccurate sensors by inte- We can dynamically predict someone’s
of the particles represents the person’s grating measurements over time. How- destination in real time as he or she moves
position, orientation, and motion veloc- ever, we could use particle filters even more about the environment, which is invalu-
Figure 4. A particle filter for location estimation using infrared and ultrasound sensors: (a) A person wearing an infrared badge
and ultrasound tags starts in the upper right corner. (b) After the person moves out of the cubicles, the samples are spread over
different locations. (c) An ultrasound sensor detects the person.
able to applications seeking to take pre- range finders provide highly accurate ID sensors. The solid and dotted lines are
dictive action. (More details on such learn- location estimates but do not identify trajectories of persons A and B, respec-
ing approaches appear elsewhere.2,10) objects. In this example, we describe tively. As they walk, the anonymous sensor
how to use Bayesian filter techniques to observes their locations frequently. Initially,
Combining anonymous and combine information from both types of their identity is unknown, but they are sep-
ID sensors sensors to get accurate location and ID arated enough to be tracked reliably using
One of the disadvantages of common information. the anonymous sensor. Until they reach
ID sensors is their limited accuracy. On Figure 6 illustrates the problem of track- ID sensor areas 3 and 4, both trajectories
the other hand, sensors such as laser ing multiple people with anonymous and have the same probability of belonging
Figure 5. A particle filter constrained to a Voronoi graph structure. The particles move along the graph’s edges: (a) The initial location
of the person is not known and the particles are spread over the entire graph. (b) After the person moves out of the cubicles, the
samples start to concentrate in the upper part of the graph. (c) An ultrasound sensor detects the person.
to either person A or B. Passing through The multitarget-tracking community tinuations—A going down and B going
the coverage of ID sensors 3 and 4 refers to this problem as the data-associ- up, or B going down and A going up.
resolves the ambiguity and determines ation problem.4 Without ID sensors, To solve the identity-estimation prob-
the IDs of both trajectories. Then, after resolving this ambiguity would be impos- lem, we use a combination of particle fil-
the paths cross, confusion exists about sible. In our scenario, we can resolve the ters and Kalman filters, closely related
the continuation of the two tracks. When ambiguity as soon as the people reach the to MHT. The key idea is to track indi-
the people leave the light-gray area, the areas covered by ID sensors 5 and 6. To vidual people using Kalman filters,
anonymous sensor can’t determine which do so, however, requires maintaining the which is possible owing to the accuracy
trajectory to associate with which person. hypotheses for both possible track con- of laser range data (see Figure 3c). A par-
ticle filter maintains multiple hypothe-
ses regarding the identities of people,
where each particle is one hypothesis for
the identity of each tracked person. Put
differently, each particle is a collection
of Kalman filters annotated with identi-
ties.11 The resulting approach can track
multiple people and their identities.
Figure 7 shows a small part of the data
set used for evaluating the approach. We
collected the complete data set with six
people, each of whom moved between
230 and 410 meters. In this challenging
sensor log, people’s paths frequently
crossed, and situations existed in which
Laser range finder up to four people were occluded to the
Ultrasound receiver laser range finders by other people. Nev-
Infrared receiver ertheless, our approach successfully esti-
mated each person’s location and identity.
W
e have codified our sensor filters as a powerful tool for state estimation in robotics. His current research projects
fusion techniques in a pro- include multirobot coordination and human activity recognition. He obtained his PhD
from the University of Bonn, Germany, in the area of state estimation in mobile robot-
ject called the Location
ics. He received an NSF CAREER award and several best paper awards at major robot-
Stack, a general framework ics and AI conferences. He is a member of the IEEE and AAAI. Contact him at the Uni-
for location-aware pervasive computing versity of Washington, Dept. of Computer Science & Engineering Campus Box 352350, Seattle, WA
98195; fox@cs.washington.edu; www.cs.washington.edu/homes/fox.
with a publicly available implementa-
tion.12 We are applying Bayes filters to Jeffrey Hightower is a doctoral candidate at the University of Washington. His re-
more complex scenarios, such as learning search interests are in employing devices, services, sensors, and interfaces so com-
puting can calmly fade into the background of daily life. Specifically, he investigates
a person’s activities from long-term sensor
design abstractions and statistical sensor fusion techniques for location sensing. He
logs. More complex estimation problems received an MS in computer science and engineering from the University of Wash-
apply structured versions of Bayes filters, ington. He is a member of the ACM and the IEEE. Contact him at jeffro@cs.washing-
ton.edu; www.cs.washington.edu/homes/jeffro.
such as dynamic Bayesian networks.13
We strongly believe that probabilistic-
filter techniques have tremendous poten- Lin Liao is a graduate student in the Department of Computer Science and Engineer-
ing at the University of Washington, Seattle. He is interested in probabilistic approaches
tial for inference problems in pervasive
to artificial intelligence. He received a MS in computer science from the University of
computing. Location estimation is just Washington, Seattle. Contact him at the University of Washington, Dept. of Computer
the beginning of connecting Bayesian Science and Engineering, Box 352350, Seattle, WA 98195; liaolin@cs.washington.edu.
reasoning systems to raw sensor data.
The full power of such techniques lies in
their ability to represent uncertainty at
Dirk Schulz is a research associate in the Department of Computer Science and En-
different levels of abstractions, thereby
gineering at the University of Washington. His main research interests are in the field
enabling truly context-aware pervasive of mobile robotics, probabilistic state estimation, object tracking, and Web-
computing systems. controlled mobile robots. He received his PhD in computer science from the Univer-
sity of Bonn in 2002. Contact him at the University of Washington, Dept. of
Computer Science and Engineering, Campus Box 352350, Seattle, WA 98195;
REFERENCES schulz@cs.washington.edu; www.cs.washington.edu/homes/schulz.