PROCEEDINGS OF THE IEEE, 2003, TO APPEAR. 1Collaborative Signal and Information Processing:An Information Directed ApproachFeng Zhao, Jie Liu, Juan Liu, Leonidas Guibas, and James ReichAbstract — This article describes information-based ap-proaches to processing and organizing spatially distributed,multi-modal sensor data in a sensor network. Energy con-strained networked sensing systems must rely on collabora-tive signal and information processing (CSIP) to dynami-cally allocate resources, maintain multiple sensing foci, andattend to new stimuli of interest, all based on task require-ments and resource constraints. Target tracking is an essen-tial capability for sensor networks and is used as a canoni-cal problem for studying information organization problemsin CSIP. After formulating a CSIP tracking problem in adistributed constrained optimization framework, the paperdescribes IDSQ and other techniques for tracking individualtargets as well as combinatorial tracking problems such ascounting targets. Results from simulations and experimen-tal implementations have demonstrated that these informa-tion based approaches are scalable and make efficient use ofscarce sensing and communication resources.Keywords— Sensor networks, distributed sensing, collabo-rative signal and information processing, information utility,constrained optimization, target tracking.Category: Collaborative Signal Information processing,Target Classification and TrackingI. Sensor Network Applications, Constraints,and ChallengesNetworked sensing offers unique advantages over tra-ditional centralized approaches. Dense networks of dis-tributed networked sensors can improve perceived signal-to-noise ratio (SNR) by decreasing average distances fromsensor to target. Increased energy efficiency in communi-cations is enabled by the multi-hop topology of the net-work [22]. Moreover, additional relevant information fromother sensors can be aggregated during this multi-hoptransmission through in-network processing [13]. But per-haps the greatest advantages of networked sensing are inimproved robustness and scalability. A decentralized sens-ing system is inherently more robust against individualsensor node or link failures, because of redundancy in thenetwork. Decentralized algorithms are also far more scal-able in practical deployment, and may be the only way toachieve the large scales needed for some applications.A sensor network is designed to perform a set of high-level information processing tasks such as detection, track-This work is supported in part by the Defense Advanced ResearchProjects Agency (DARPA) under contract number F30602-00-C-0139through the Sensor Information Technology Program. The views andconclusions contained herein are those of the authors and should notbe interpreted as representing the official policies, either expressed orimplied, of the Defense Advanced Research Projects Agency or theUS Government.Feng Zhao, Jie Liu, Juan Liu, and James Reich are with Palo AltoResearch Center (PARC), 3333 Coyote Hill Road, Palo Alto, CA94304, USA (email: {zhao,jieliu,jjliu,jreich}@parc.com). LeonidasGuibas is with Computer Science Department, Stanford University,Stanford, CA 94305, USA (email: [email protected]).ing, or classification. Measures of performance for thesetasks are well defined, including detection, false alarms ormisses, classification errors, and track quality. Commercialand military applications include environmental monitor-ing (e.g., traffic, habitat, security), industrial sensing anddiagnostics (e.g., factory, appliances), infrastructure pro-tection (e.g., power grid, water distributions), and battle-field awareness (e.g., multi-target tracking).Unlike a centralized system, however, a sensor networkis subject to a unique set of resource constraints such aslimited on-board battery power and limited network com-munication bandwidth. In a typical sensor network, eachsensor node operates untethered and has a microprocessorand limited amount of memory for signal processing andtask scheduling. Each node also is equipped with one ormore of acoustic microphone arrays, video or still cameras,IR, seismic, or magnetic sensing devices. Each sensor nodecommunicates wirelessly with a small number of local nodeswithin the radio communication range.The current generation of wireless sensor hardwareranges from shoe-box sized Sensoria WINS NG sensors [20]with an SH-4 microprocessor to matchbox sized Berkeleymotes with an 8-bit microcontroller [12]. It is well-knownthat communicating one bit over the wireless medium con-sumes far more energy than processing the bit. For theSensoria sensors and Berkeley motes, the ratio of energyconsumption for communication and computation is in therange of 1,000–10,000. Despite the advances in silicon fab-rication technologies, wireless communication will continueto dominate the energy consumption of embedded net-worked systems for the foreseeable future [8]. Thus, min-imizing the amount and range of communication as muchas possible, for example, through local collaboration, datacompression, or invoking only the nodes that are relevantto a given task, can significantly prolong the lifetime of asensor network and leave nodes free to support multi-useroperations.Traditional signal processing approaches have focused onoptimizing estimation quality for a fixed set of available re-sources. However, for power-limited and multi-user decen-tralized systems, it becomes critical to carefully select theembedded sensor nodes that participate in the sensor col-laboration, balancing the information contribution of eachagainst its resource consumption or potential utility forother users. This approach is especially important in densenetworks, where many measurements may be highly redun-dant, and communication throughput severely limited. Weuse the term “collaborative signal and information process-ing” (CSIP) to refer to signal and information processingproblems dominated by this issue of selecting embeddedPROCEEDINGS OF THE IEEE, 2003, TO APPEAR. 2sensors to participate in estimation.This paper uses tracking as a representative problem toexpose the key issues for CSIP — how to dynamically de-termine what needs to be sensed, who should sense, how of-ten the information must be communicated, and to whom.The rest of the paper is organized as follows. Section II willintroduce the tracking problem and present a set of designconsiderations for CSIP applications.