WP5 - 5:00 - EECS at UC Berkeley

0 downloads 0 Views 2MB Size Report
confederacy of identical autonomous cooperating processes. The general ... At any one measurement time, each sensor measures ... But at one seasurement time, it usually estimates the ... times produced by the first algorithm so complicates.
WP5 - 5:00 DISTRIBUTED ESTIMTION IN THE NIT/LL DSN TESTBED J. R. Delaney, R. T. Lacoss and P. E. Green HIT Lincoln Laboratory P.O. Box 73 Lexington, MA 02173-0073

1. Introduction and Sumary A DSN is a surveillance and tracking system employing many geographically dispersed sensor/processor nodes connected by a computer cosunications network and implemented as a confederacy of identical autonomous cooperating processes. The general properties of a DSN are being investigated through the development and exercise of a test-bed system. The detection and tracking of low-flying aircraft using simple acoustic sensors has been selected as a specific problem to be addressed. Each node in the test-bed consists of an array of microphones as the sensor, several small computers for data processing, and a digital radio for cosunicating information between the nodes. Use of microphone arrays as sensors lmits each node to measuring target azimuths; nodes ust exchange target azimuth information in order to track target positions.

A DSN need not be as simple or as hoogeneous as the test-bed. It can mix sensor types such as radar and passive IR, putting one type or both at any node. Data processors need not be colocated with all sensors and vice versa. Data processing capacity can vary from node to node and comunications capacity can vary from link to link. The counications network ca mix broadcast radio, point-to-point radio, and wire. But it is our opinion that all of the significant DSNrelated tech al issues which must be addressed in developing such comlex systems must also be addressed in developing the test-bed system and that the relative simplicity of the test-bed helps focus attention on those DSN-related issues. This paper describes the acoustic tracking algorithms currently used in the MIT Lincoln Laboratory (MIT/LL) Distributed Sensor Networks (DSN) test-bed. It discusses the original motivation for inclusion of various features in those algorithms and the lessons learned about those features through experimentation with real and simulated data. Plans f or modifications to the detection and tracking algorithm 're briefly sketched.

Acoustic propagation introduces delays in azimuth measurements which complicate the merger of target azimuth information into target position estimates. At any one measurement time, each sensor measures target azimuths that correspond to different times in the past, depending on target-to-sensor range. A node cannot estimate target position from just the most recent measurements because they correspond to different target positions. Each node must maintain a history of past sensor measurements in order to track targets. Two algorithms have been developed for estimating target positions from histories of sensor measurements produced by two or more nodes. One algorithm has the advantage that it introduces no more delay into the target detection and tracking process than is forced by acoustic propagation. But at one seasurement time, it usually estimates the positions for different targets at

different times. Over a regular sequence of measurement times, it usually estimates the position of the same target at irregularly spaced times. The other algorithm forms estimates at each measurement time with a specified delay which is uniform for all targets. The product of the delay and the speed of sound is the mim range at which targets can be tracked, so thedelay is usually set equal to the sensor detection range divided by the speed of sound. On the surface, the generally larger delay of the latter algorithm makes it appear undesirable. twever, the irregularity of estimation times produced by the first algorithm so complicates further data processing as to make it even less attractive. The microphone arrays used in the test-bed have a detection range on the order of 5 to 10 km and the radios have a similar range. Thus, no one node can measure a target's azimuth for very long. This difficulty can be overcome if each node's azimuth data is distributed to every other node. But to do so would be difficult in a large DSN, producing heavy cosunications traffic and requiring much redundant data processing in each node. The least distribution practical, broadcast of azimuth data only to nodes in direct radio contact, is done in the test-bed so as to hold down conunications traffic.

Only azimuth data is now exchanged by the test-bed nodes. Because this data is exchanged only between nodes in direct radio contact and because the radio range is comparable to the sensor range, each node has sufficient information to track the positions of targets within its sensor coverage. This restriction impacts target position tracking by individual nodes only in the area of track initiation. As a target moves through the DSN, each node must acquire the target as it enters the node's sensor coverage and initiate anew a position track for the target despite the likely existence of position tracks for that target in other nodes. This restriction also has a major impact on the surveillance function of the DSN as a whole because individual nodes have a myopic view of the targets within the coverage of all the DSN's sensors. Intuitively, exchanging position tracks as well inuth data between nodes in direct radio contact should allow target position tracks to be handed-over

as

instead of re-initiated. And more extensive comunication of position tracks should allow formation of a complete surveillance picture. But it.is not yet clear how to combine, in a statistically valid manner, position tracks formed in different nodes using semsor data which is not all exchanged. Recent research has examined related, simpler problems, and suggests that this difficulty is surmountable. However, the applicability of the research results needs further study.

305 Authorized licensed use limited to: Univ of Calif Berkeley. Downloaded on August 20, 2009 at 15:01 from IEEE Xplore. Restrictions apply.

Figure 2 expands the tracking block in Figure 1 and includes, for reference, the micropbone array md the signal processor. It show all of the data flow in the detection and tracking system, including those between nodes. The data flows are complex but, as will be seen, the complex flows are necessary. The data from each microphone array is procesed very two seconds using adaptive, nonllinear filtering1- to detect local maxim of the incident sound power (averaged over the preceding two seconds) in frequency and azimuth angle. A measuremnt of each detection's frequency, azimuth angle, and average sound pressure level is produced in the process.

IT/LL DSN Test-Bed

2.

Figure 1 showe a simple block diagram of one D6M test-bed node as well as photographe of a microphone array and a mobile node vehicle carrying the associated computers, radio, and their power supply. Atmospheric conditions limit the M_xiim range at which typical targets can be detected to betwee 5 and 10 km. The test-bed radios, currently under construction, will typically be limited to a similar range by line-of-sight propagation. Thus, each node cosunicates directly with those nodes having overlapping sensor coverage and only those nodes. The radios are also designed to measure the range between iodes, allowing the test-bed to estimate the relative locatious of its nodes as indicated in the figure. Until the packet radios are available, broadcasts are simulated using wire comunications.

D/WTRBUTED SEPNOM NETWORKS * 'hm-o_

%^

*

..

f t I.

.

--

PMET

RAWI

RA

,-r

.";

IC5ยง.

stkV

0 Y

~~'' CONCEPT

j.

r it

i-

TEST NODE t -

-.

Fig. 1.

The MIT/LL DSN Test-Bed.

306 Authorized licensed use limited to: Univ of Calif Berkeley. Downloaded on August 20, 2009 at 15:01 from IEEE Xplore. Restrictions apply.

FROM OTHER NODES

TO

UPATD AIU __ ~~~~ESTIMATE-S , __ AZIMUTH TRACK SOAE AZIMUTH/AZIMUTH RAT

SIGNAL PROCBESSOR

OTHER MODES

AZ1JAUTH RCE

A6nuh Prsuv

SAMSLED __

VOLT GES ARIRAYX

Fig. 2.

Functional Description of The NIT/aL DSN Test-Bed Tracking System.

The detections and measurements are made over a four to five octave frequency band. Such broad bandwidth allows detection of multiple harmonic emissions by a single target. Sensor data conditioning is done to reduce tracking computational load by clustering together detections which could plausibly have been caused by a single target and by discarding detections which are relatively weak in sound pressure level. Each cluster is characterized by the average azimuth angle of the component detections (weighted by sound pressure level) and by the total sound pressure level. The average azimuth angles are treated as target measurements thereafter. Acoustic propagation variability makes incident sound pressure level a second-order measure of target range; the sensors are primarily azimuth-only measurement devices. Thus, the full target state is not observable using a single sensor and no node can accurately estimate it without information from other nodes; the nodes must share measurement information to form an estimate of the full target state.

The tracker uses a two state (azimuth angle and azimuth angle rate) a6- tracking algorithm for estimation and prediction. Data association is done quite simply. A measurement is associated with an azimuth track and used to update the state estimate if the azisuth angle lies vithin a window about the filter's azimuth angle prediction for the measuresent time. Only one association is allowed per track or per detection. Should a measurement fail to. associate with any existing track, it is used to initiate a new track. A newly initiated track is terminated if no measurement is associated with it at the next measurement time; any other track is allowed only one missed data association in a row. Full azimuth track state estimates are maintained within each node for the targets sensed locally but only the the azimuth component is broadcast to other nodes. The azimuth components computed locally are also added to the node's azimuth history data base along with all azimuth components received from other nodes. Each azimuth track created by a node is given a unique (within the node) identifier which is broadcast with each azimuth component. This tagging allows azimuth components broadcast at different times but based on the sw azimuth track to be associated. Figure 3 sketches the organization of the azimuth history data base. It is tree-structured, with the data sorted by originating node and azimuth track.

Only the data in the azimuth history data base is used in further processing. The effect is to treat the each node's sensor, signal processor, sensor conditioning, and azimuth tracker as a virtual sensor with the measurement properties of the whole chain. Each node can be thought of as connected to a number of such virtual sensors, one at its location and the others at the locations of those nodes with which it has direct radio contact. This artifice helps modularize the tracking system, decoupling details of the sensor, signal processor, et cetera, from the remainder of the tracking system. The remainder of the tracking system mist form estimates of target dynamic state (position and velocity) from these virtual measurements of acoustic azimuth. The process is done in two steps at each node:

1)

Estimate target positions from acoustic azimuth data in the azimuth history data base, and

2)

Estimate target dynamic states from the estimated target positions.

Low-flying aircraft can travel at an appreciable

fraction of the speed of sound and even exceed it in some cases. As a result, the sensors measure target azimuth angle not at the measurement times but at the times when the measured sounds were emitted. Target azimuth angles corresponding to present target positions are not observable. The lack of observability prevents the direct application of familiar tracking techniques, including the Kalmn filter, which update target state estimates using current measurement data only. Further complications are the nonlinearity of the measurement process and its effective noncausality; a measurement by one seasor may be made later than a measurement by another sensor but my contain information about the target state at an earlier time in the target's fram of reference.

AZIMUTH HISTORY DATA BASE

NODES

AZIMUTH TRACK IDENTIFIERS

Target azimuth data must be accumulated over time in order to form target position tracks. The first step in this accumulation is target azimuth tracking. The azimuth tracker in each node is quite conventional, with theexception that it does not estimate the unobservable present target azimuth angle but rather the observable 'acoustic" azimuth angle.

Fig. 3.

Azimuth History Data Base Organization.

307 Authorized licensed use limited to: Univ of Calif Berkeley. Downloaded on August 20, 2009 at 15:01 from IEEE Xplore. Restrictions apply.

POSITION TRACK DATA BASE

Azimuth histories can be combined in two ways to produce position estimates. The current tracking system uses the reflection algorithmx84,5x9. The

algorithm takes as inputs two azimuth histories originating from differentnodes and the locations of those nodes. If the two azimuth histories could plausibly be associated with a single target, the algorithm estimates the position of that target for the emission time of the sound just detected at the node closer to the target. A position estimate produced by the algorithm can be regarded as a "measurement" of the target's position, delayed in availability depending on the target's range from the sensor. The delay is the least achievable with any position estimation algorithm; it equals the shortest propagation delay in the current azimuth measurement

NODE PAIRS

AZIMUTH TRACK IDENTIFICATION PAIRS

(to, x(to), y(t)) ESTIMATE Fig. 4.

data.

In each node, the reflection algorithm is applied to all legitimate combinations of azimuth histories. The resulting position estimates for different targets are for different times. The times csn range from the most recent measurement time (for targets overflying sensors) to 30 seconds earlier (for targets 10 km. from the nearest sensor). Because target-to-node range varies with time, so does the delay for a single target's position estimates atsequertial measurement times, even if the estimates are based on the same pair of azimuth histories (same nodes and track identifiers) at sequential measurement times. The variable times associated with target position estimates produced by the reflection algorithm complicate the process of building position tracks froe position estimates. Consider a target which is tracked simultaneously by three test-bed nodes and assume that each node is within radio range of the other two. Then each node contains three azimuth histories corresponding to the target. Applying the reflection algorithm to each pair of histories would typically yield three different position estimates, each for a different time. It is very difficult to recognize that all of these position estimates correspond to the same target, even if the azimuth histories are completely accurate and if the reflection algorithm introduces no inaccuracies. For this reason, position tracks are currently formed only from position estimates based on particular pairs of azimuth histories, i.e., for the same pair of nodes and azimuth tracks within those nodes. To facilitate thisisolation of position tracks, they are kept in a tree-structured data base organized like the azimuth history data base. In this case, entries are sorted first by the originating pair of nodes and then by the pair of identifiers of the originating azimuth tracks (see Figure 4). Position track data base entries consist of a time and estimates of target easting, northing, east velocity, and north velocity at that time, plus some auxiliary

track.

Orgaization.

Neither target position estimates nor position tracks are exchanged between nodes in the current test-bed. If the nodes are all directly connected by radio to each other, such an exchange would be unnecessary. Each node would (in the absence of co=unications failures) have identical azimuth history data bases and produce identical target position estimates and position tracks. But complete connectivity is not practical in large DSNs; thus, azimuth history data bases typically differ from node to node and so must the position estimates and position tracks based upon the azimuth history data. Each node then has an incomplete picture of the targets in the DSN's coverage. This point is illustrated in the next section. Exchanging target position estimates or position tracks could provide nodes with additional information, but consider an extreme example of what can go wrong if target state estimates are not properly combined: A local state estimate a is created in data processor A from a single measurement by one sensor. That estimate is shared with data processor B, which .uses it to create an identical local state estimate, 8. Data processor B shares its local state estimates with data processor A, including B. Data processor A associates local state estimates a and B with the same target and combines them to produce local state estimate y with half the variance of a and 0. This process could ultimately lead to the existence in data processors A and B of a local state estimate w of infinitesimal variance based on a single sensor

measurement.

Given the complexity of this issue, the decision wms made to limit the initial version of the test-bed to exchanging only the azimuth components of azimuth tracks and only between nodes in direct radio contact. Later versions of the test-bed will experiment with more extensive data exchanges.

information.

A position track is updated when a new position estimate is produced for the pair of azimuth tracks histories upon which that track is based. A new position track is created if no entry exists in the data base corresponding to an azimuth track. pair which passed the reflection algorithm's test. Aa-$ tracking algorithm is again used for prediction and estimation. Azimuth histories are not completely accurate and the reflection algorithm can amplify those inaccuracies. The resulting position estimates can be inaccurate not only in position but also in time. For this reason, position estimate times are smoothed before the a-B algorithm is applied. The Auxiliary inf ormation in each position track data base entry is the state of the smoother for that position

Position Track -Data Base

3. Tracking Performance

All of the results shown in this section are for Four test--bed nodes were one field used to record the passage of a UH-1 helicopter west-to east along the flight path illustrated in Figure 5 at a ground speed of roughly 65 knots and at

experiment;i6.

roughly 1000 feet above ground level. The letters F, H, J, and L indicate the locations of the four nodes. The circles indicate checkpoints used by the helicopter pilot to maintain his f light path and by observers to time the helicopter's passage. 308

Authorized licensed use limited to: Univ of Calif Berkeley. Downloaded on August 20, 2009 at 15:01 from IEEE Xplore. Restrictions apply.

The curve marked "TRACK" is the track of the UH-1 helicopter, which reached its point of closest approach at roughly 240 seconds and was roughly 600 meters from the microphone array at that time. Two other sound sources were tracked: a fixed-wing aircratt with a varying azimuth angle and a bulldozer with an essentially constant azimuth angle (due south). The "speckles" on the plot are signal processing artifacts or very intermittent sound sources. Such artifacts or sources are detected uniformly over time, but are suppressed by the sensor data conditioning process early in the plotted data because of a relatively loud target near the sensor.

Fig. 5.

UH-1 Flight Path for November 1981

Experiment.

The runways overflown near checkpoints 3 and 4 are Hanscom Field, a busy military and civilian airfield. Under checkpoint 6 is Route 128, a heavily used eight-lane superhighway. Another major highway, Route 2, lies just south of the area shown on the map. Normal activity at Hansccm field provided additional aircraft in sensor coverage during the experiment and normal activity on Route 128 provided an extended and irregular interfering soundsource. Construction vehicles and stationary mechanical equipment operating near the microphone arrays also provided acoustic interference.

Figure 6 is an azimuth-time-intensity plot for the output of the signal processor at node F. The plot only includes those measurements at each time having a sound pressure level within 10 dB. of the maximu value at that time. The curves overlying the measured values are the sequences of azimuth components of the azimuth tracks produced when processing the data. Each continuous curve corresponds to one track. Because of breaks in the measurement data, several distinct tracks (having distinct track identifiers) are caused by each sound

Figure 7 shows the position tracks formed using updated azimuth estimates calculated in node F (shown in the previous figure) and in node H. Each cross is the position component of a newly updated position track. Thin lines connect sequential position components of individual position tracks. The line marked "TRACK" is again the track of the UH-I helicopter. That track is crossed by another, that of the fixed-wing aircraft. The other tracks are of intermittent sound sources or are processing artifacts, e.g., "ghost" tracks based on erroneous pairings of azimuth track histories which manage to pass the reflection algorithm test. The short, dense track radiating from node H is a "ghost". Ghost tracks can often be recognized because they trace out physically unreasonable trajectories, beginning or ending at a node or involving unrealistic accelerations.

Recording of the measurement data allowed experimentation with different connectivities between the nodes. Figure 8 shows the position tracks formed in node J when it received azimuth data broadcast only by nodes H and L. The figure includes distinct but overlapping tracks of the UE-I helicopter, each derived from a different pair of azimuth histories. Note that node J has an incomplete surveillance picture. It does not include the track of the fixed-wing aircraft which appears in Figure 7. The latter track was based in part on azimuth information sensed at node F; information unavailable to node J in this case.

source. 360

2500 2000

TRACK

TRACK

111U1 rEASTBOUND

UHI EASTBOUND; 1000

1

270 El) va

M

0

UD LU

0)

.

-1000 10 so

40

428 '

-2000

1i2 Fig. 6.

200

300 SECONDS

-2500-I

402 428

-4000

-2000

0

1000

M

Azimuth Angle Measurements and Azimuth Tracks for Node F.

Fig. 7.

Position Tracks formed at Node F from hzimth Histories originating at Nodes F and R

309 Authorized licensed use limited to: Univ of Calif Berkeley. Downloaded on August 20, 2009 at 15:01 from IEEE Xplore. Restrictions apply.

Entries in the position track data base would not need to be sorted by originating azimuth history pairs and more usual data association could be done. Smoothing of the position estimate times would be unnecessary, eliminating this potential source of error. Experiments using the possible position algorithm for target position estimation, and the simplified position track data base and tracking algorithm it allows, are planned to evaluate the trade-off s between timeliness and complexity of the two algorithms.

*si(

The limited sharing of information between nodes in the current test-bed does not interfere with position tracking by individual nodes. But it prevents nodes from "handing over" target position tracks as the targets pass through the DSN's sensor coverage and can cause each node to have a myopic view of the targets in the DSN's overall sensor coverage as illustrated in the previous section. Results of recent research9-lu into related problems suggest that it should be possible to form a complete surveillance picture in each data processor if each one transmits locally computed target state estimates to other data processors and if each uses the state estimates it receives as well as available raw sensor data to update its local target state estimates. But the research results demonstrated only asymptotic convergence of estimates of an invariant quantity. We must examine this work carefully to extract clues as to the proper organization of algorithms for our more complex situation. At the very least, we would like to develop an ad hoc algorithm for combining state estimates for the purposes of position track hand-over that exhibits minimal pathological behavior. The need to demonstrate a system which performs satisfactorily in a realistic environment will continue to drive the development of the Lincoln DSN test-bed Laboratory tracking system. In parallel with the development of new tracking algorithms, data on actual sound sources in stressing will be collected as environmnts such crossing targets on known trajectories. Regular experisentation with such data of the system on significant focuses point for theoretical providing a starting problems,development investigations and a timely test of any ad hoc or

-1000 -2000 -2500 -4000

Fig. 8.

-2000

-0

1000

Position Tracks formed at Node J from Azimuth, Histories originating at Nodes H, J and L.

4. Lessons and Plans Use of the reflection algorithm for target position estimation has the advantage of producing the most up-to-date position estimates possilble. But it has the disadvantage of producing positi on estimates with time-varying delays, making it diff:lcult to recognize position estimates associated iwith the target. A consequence is the overlappingg TI-i position tracks in Figure 8. Location egstimate time smoothing was required in the position tiracking algorithm as a consequence of inaccuraciias in the time-varying delays. Experiments using isimulated measurement data with no azimuth measurei sent inaccuracies has revealed that this smoolthing process by itself produces some inaccuracies in 1the position tracks. These observations have led us 1 whether the advantage of the reflection galgoristios worth the disadvantage.

avprqximate aspects

Another position estimation algoriti im, the possible position algorithmI48, takes tha same inputs as the reflection algorithm and produces position estimates with a fixed delay. Targets at tranges great enough for acoustic propagation time to X axceed that delay are ignored by this algorithm even if they are detected. So the delay is usually chose: 1 to equal the propagation time for sound from a target at the maximum detection range. This delay woul tdbe 3 seconds for the test-bed. For a target Niery near a node, the possible position algorithm cat ve a delay nearly as large between detection of the target by that node's sensor and the creation of a corresponding position estimate. Lesser delays would c3ncur for targets further from nodes. Since targetts are typically first detected away from all nc)des and are well in track by the time they are close to any node, this disadvantage is probably less signilfFicant tha it appears on the surface.

of the

system.

Rferences i.

2.

J. Capon, "High Resolution Frequency-Wavenusber

Spectrum Analysis", Proceedings of the IEEE, v. 57, n. 8, pp. 1408-1418 (August 1969). "Distributed Sensor Networks

Institute of Technology

September 1979) A086800.

3.

Semiannual

, Lexington, MA (30

"Distributed Sensor Networks Semiannual Technical Sumary", Lincoln Laboratory, Massachusetts

Institute of Technology , Lexington, September 1980) A103045.

4.

The availability of position estimat'es at regular and common times should allow significant simplification of the position tracker. Location estimates plausibly caused by the same tairget could be clustered together in the same manner as are azimuth measurements produced by the the signal p'rocessor.

MA

(30

R.P. Hughes, "A Distributed Multiobject Tracking Algorithm for Passive Sensor Networks", Master's Thesis, Massachusetts Institute of

Technology (June

5*

Technical

Sumary", Lincoln Laboratory, Massachusetts

1980).

J.R. Delaney, "Location Algorithm Sensitivities", in preparation as Technical Note, Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, MA.

310 Authorized licensed use limited to: Univ of Calif Berkeley. Downloaded on August 20, 2009 at 15:01 from IEEE Xplore. Restrictions apply.

References (cont'd) 6.

7.

P.E. Green, "DSN Test-Bed Tour and Demonstration', Distributed Sensor Networks Workshop, Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, MA, pp. 163-174 (January 1982) A118868.

"Distributed Sensor Networks Semiannual Technical

Summary", Lincoln Laboratory, Massachusetts Institute of Technology , Lexington, MA (31 March 1978) A060685.

8.

R.L. Walton, "Multisite Acoustic Location", Distributed Sensor Networks Workshop, Computer Science Departsent, Caraegie-Mellon University, Pittsburg, PA, pp. 63-67 (December 1978).

9.

V. Borkar. and P.P. Varaiya, "Asymptotic Agreement in Distributed Estimation", IEEE Transactions on Automatic Control, v. AC-27, n. 3, pp. 650-655

(June 1982).

10.

J.N Tsitsiklis and M. Athans, "Convergence and Asymptotic Agreement in Distributed Decision Problems", Proceedings of the 5th MIT/ONR Workshop on Distributed Information and Decision Systems, Monterey, CA (August 1982).

This work was sponsored by the Defense Advanced Research Projects Agency. The U.S. Government assumes no responsibility for -the information presented.

311 Authorized licensed use limited to: Univ of Calif Berkeley. Downloaded on August 20, 2009 at 15:01 from IEEE Xplore. Restrictions apply.

P~~~~'2ACK 2ET ACOSTIC

,.rRAY s RADIO~~~~~~~

_WS~ ~

CON~~~~~~~~~~~~ 1I1 COMUNIAT

ARA MICROPHONE~~~~~~~

AN

SLLOCATICN MO-M6000)}Xlb|

DATA~ COLLE

~~~MBL TES NODE

11

O

s 12011

TET OD Fig. 1. The MT/LL DSN Test-Bed.;

W