MIT Lincoln Laboratory - Bedford, MA

August, 1984 - May, 1985

DISTRIBUTED SENSOR NETWORKS 1. EXECUTIVE SUMMARY The overall DARPA Distributed Sensor Networks (DSN) program involved several research organizations and was aimed at developing distributed target surveillance and tracking methods for systems employing multiple spatially distributed sensors and processing resources. Such systems would be made up of sensors, data bases, and processors distributed throughout an area and interconnected by an appropriate digital data communication system. The working hypothesis of the program was that through netting and distributed processing, the information from many sensors could be combined to yield effective surveillance systems. The overall con- cept called for a mix of sensor types as well as geographically distributed sensors. The Lincoln Laboratory program emphasized the development of an experimental test bed to demonstrate DSN concepts. Surveillance and tracking of subsonic low-flying aircraft using ground based acoustic and imaging sensors was selected as the focus of the system development and demonstrations. Small arrays of microphones providing directional information were employed as acoustic sensors, and visible wavelength TV cameras were used as imaging sensors in the test bed. A primary objective of the Lincoln Laboratory DSN effort was to prove the feasibility of distributed surveillance systems by demonstrating real-time distributed tracking of low flying air- craft. This was accomplished. The real-time acoustic and TV tracking experiments described in Section 9 constitute the primary feasibility demonstration. They involved real-time helicopter tracking using an experimental system composed of six sensor/processor nodes deployed in the vicinity of Hanscom Field. The major part of the effort was the development and test-bed implementation of distributed algorithms to demonstrate distributed tracking. Most of this report deals with these topics. Algo- rithms and test-bed systems are described and discussed in detail. Experimental results. involv- ing simulated as well as real data, are included to illustrate concepts and validate algorithms. Other topics that are covered include network self-location and multi-site data integration. The test-bed tracking system is implemented using autonomous cooperative processes operating at each of the nodes. Once started in operation, each tracking process interacts with those at other nodes through discrete data messages, with no external control. Each node per- forms tracking with whatever data are available to it: data derived from attached sensors and data obtained through messages from other nodes. This structure has resulted in a tracking system that easily adapts to changes in the number and type of sensor/processor nodes in the network. When a node fails the only impact on the overall system is the degradation in tracking performance due to the loss of the sensor at the node. Real-time distributed tracking demonstrations required the development of suitable algo- rithms. Tracking algorithms were developed to: (1) perform sensor independent tracking; (2) track aircraft with small microphone arrays; and (3) perform tracking with a mix of sensor types, including acoustic and TV sensors. The DSN tracking algorithms depend upon well understood estimation methods. They can easily be adapted to accommodate new sensor types by the addition of sensor specific Kalman Filters or Extended Kalman Filters for each new type. Also, they provide estimates of present target positions despite acoustic propagation delays and automatically utilize the data from any number of DSN nodes. These algorithms replace less general acoustic tracking algorithms that were developed during the early stages of the DSN project. The approach for the earlier algo- rithms was to minimize assumptions about target dynamics. This resulted in time-delayed posi- tion estimates, ineffective use of the data from the sites nearest to the target, and provided no obvious way to combine acoustic with non-acoustic tracks. The initial algorithms also provided no theoretical model for utilizing data. The new algorithms solved these problems by modeling target dynamics and using the models as the basis for tracking in terms of present target position, independent of acoustic propagation delays. For tracking purposes the targets are now modeled as constant velocity objects subject to random accelerations. There are several tracking algorithm components. The azimuth tracking algorithm is a Kalman Filter that processes acoustic azimuth measurements into azimuth tracks. The filter assumes constant angular motion with random angular acceleration and forms azimuth tracks without reference to the geographical position of the target. The track initiation algorithm forms position tracks using pairs of azimuth tracks from two different nodes. Extended Kalman Filters (EKF) are used to update position tracks with acoustic azimuth measurements or line-of-sight azimuth measurements obtained from TV cameras. All algorithms compensate for acoustic delays when needed so that tracking is in terms of present target position estimates. 2 The Extended Kalman Filters in a node process only measurements derived from the sen- sors attached directly to the node. Sensor information from other nodes is indirectly utilized by forming optimal Bayesian combinations of local position tracks with position tracks received in messages from the other nodes. This combining is independent of the sensor types since it deals only with target state and error covariance estimates. The combining algorithm is the mechanism we have used to decompose and distribute the tracking function so each node is concerned only with its own attached sensors. Other tasks handled by the tracking algorithms include data association, track association, and track maintenance. We elected to implement these functions as simply as possible. In addi- tion we elected to allow each measurement to be associated with only one track. One result is that alternative interpretations of the measurement data are not considered. These are areas in which significant improvements could be made to the algorithms. Data association is the assignment of new sensor measurements to tracks. This is done by comparing new measurements with tracks and associating a measurement with the first track that yields a good match. The match test is statistical and takes measurement errors and the track errors into account. Track association identifies a track estimate from one node with the correctly corresponding track estimate at another node based on track identifiers that are assigned when tracks are ini- tiated. During normal system operation tracks are initiated by one or two nodes. For the two node case, the track identification rules are such that both assign the same identifier. Subse- quently, other network nodes are alerted to approaching targets and given the correct identifier. Track maintenance involves deciding when to add or delete a track from the track data base in a node. Tracks are deleted when their estimated errors exceed a threshold. A track can persist (coast) during periods when no new measurements are associated with the track. Without new measurements the estimated errors increase, eventually becoming large enough to cause the track to be deleted. A track generated by the track initiation algorithm is added to the data base when it satisfies a number of conditions based on its estimated error and the geometry of the situation. The last important issue for the distributed tracking algorit'ams is what information to exchange between nodes and when to broadcast it. Unprocessed sensor measurements are never 3 broadcast. Azimuth tracks are broadcast when they have not been associated with a position track. This makes them available for position track initiation. Position tracks are broadcast when their estimated errors are small enough and several other conditions are met. Some conditions involve where the estimated track is located relative to the node and to neighboring nodes that are expected to receive the broadcast. The rate at which updated position tracks are broadcast may also be reduced to limit the communication load of the network. All aspects of tracking algo- rithms are treated in detail in Section 6. Demonstration of real-time tracking required the development of acoustic signal processing algorithms to detect targets and determine directions. A wideband array processing algorithm was developed for this purpose and was implemented and used for real time tracking demonstra- tions. This algorithm is based upon a new approach to direction finding with arrays. It depends upon the following facts. First, the spatial correlation function of a plane wave is like a long mountain range with the ridge line perpendicular to the direction from which the signal is arriv- ing. This fact does not depend upon the temporal bandwidth of the plane wave. But the two dimensional spatial Fourier transform of a mountain range is another long mountain range, but this time passing through the origin of the two dimensional spatial frequency space and rotated by ninety degrees to be oriented in the direction of the arriving signal. The mountain ranges for independent plane waves arriving from different directions are simply added together. The wide- band algorithm looks for these mountain ranges radiating from the origin. Their directions are target directions and their heights are proportional to signal strength. The wideband algorithm replaced a more conventional one that was developed and used for initial DSN experimentation. The new algorithm provided more reliable detections and direction finding while requiring fewer computations than the original. Acoustic detection and direction finding algorithms are discussed in Section 4. Whereas acoustic arrays are appropriate for surveillance and track initiation as well as for tracking, the TV sensors have limited fields of view and are more appropriate for improving tracks than for initiating new ones. The TV cameras were developed to function as autonomous cued measurement resources. They operate as follows. The TV subsystem receives position track estimates and calculates which target it is most likely to be able to detect. It then slews to where that target is predicted to be and captures two frames of data which are used to attempt to 4 detect the target on the basis of motion in the image frame. If there is a detection, then, using the known orientation of the camera and the position of the detection in the image frames, the target azimuth is estimated and sent to the distributed tracking system which integrates it with all other measurements. This is a closed loop system in which the TV uses target position tracks to select and find targets and produce measurements that subsequently improve tracks. The algorithms used by TV subsystems to select targets, control the TV subsystem, detect targets and provide azimuth measurements to the tracker are described in Section 5. A major part of the Lincoln Laboratory DSN effort was the development of the test-bed sys- tem. It contains eleven nodes interconnected by Ethemets and radio communication links. Every node contains from one to three single board microcomputers that perform tracking and commu- nication functions. Six of the nodes also contain acoustic subsystems consisting of a small microphone array and the electronics required to collect and process acoustic data in real time. Some nodes are installed in vehicles for field deployment. Two of the nodes are dedicated to TV functions. The test bed also includes microwave communication equipment that allows the test bed to be split into three physically separated parts, an essential capability for experiments involving mobile nodes separated by more than about a kilometer. Two UNIX workstations and a VAX computer constitute the remainer of the system. They are employed for software develop- ment, experiment control, and other analysis functions. The test bed is a large hardware and software system that required substantial effort to develop and use. Section 3 describes the hardware and the system software. Section 7 describes the application level control and tracking software. Section 8 provides more details about the communication system. Sections 10 and 11 address the software development process and les- sons that were learned about the development of such complex distributed systems. We have attempted in these last two sections to distill and document information that we believe may be particularly useful to future implementors of Distributed Surveillance Systems. Sections 12, 13, and 14 cover problems other than tracking and test-bed development and use. Section 12 treats the problem of how to estimate the network node locations using range measurements between pairs of nodes. A distributed self-location algorithm is described along with sample results obtained using simulated range measurements. Autonomous portions of the 5 algorithm operate at each node. Each node estimates its position relative to other nearby nodes. The algorithm is iterative and continues to run as long as there are changes in the estimates. The range measurements needed by the algorithm could be obtained using radios to measure the tran- sit time of electromagnetic signals between the nodes. Radio features and protocols to measure ranges, even with unsynchronized nodal clocks, are described. Section 13 treats multi-site data collection: how to collect tracks from many nodes in a net- work and how to resolve differences between tracks provided by different nodes for the same target. The track combining algorithm developed for tracking can also be used for combining similar tracks during multi-site data collection; but this solves only part of the multi-site data collection problem. Section 13 discusses other aspects of the problem, especially the close cou- pling between communication and multi-site data collection issues. Section 14 summarizes the results of a preliminary investigation of the application of Artificial Intelligence approaches to some acoustic data interpretation problems in an acoustic DSN. The last Section discusses remaining DSN research and development issues. 6

Copyright © 2026 Mike Axel.