Book Review

Handbook of Multisensor Data Fusion

Edited by David Hall and James Llinas CRC Press, 2001

A book which includes dirty secrets in multisensor data fusion, the unscented filter and the Bayesian iceberg cannot be all bad. This handbook consists of 26 chapters, many of which are excellent, including well-written solid quantitative state-of-the-art surveys by Bar-Shalom, Poore, Uhlmann, Julier, Stone, Mahler, and Kinibarajan, among others, The chapters in this handbook range from tutorial to cutting-edge research, so there is something for everyone. Mary Nichols gives a broad and rapid survey of military multisensor data fusion systems in Chapter 22. A plethora of websites about data fusion and related subjects is included at the end of this handbook. The most ambitious and interesting chapter, by far, is on Random Set Theoryfor Target Tracking and Identification, by Ron Mahler. Random set theory is clearly a good idea for data fusion, owing to the uncertainty in the number of targets and the uncertain origin of measurements in many practical applications. I n fact, random sets are essentially required, implicitly or explicitly, to describe the problem of tracking in a dense multiple target environment. Mahlers chapter is edifying but sometimes obscure. In particular, what Mahler calls turn-the-crank formulas require derivatives and integrals with respect to sets: these are not the bread and butter of normal engineers, and hence, a few simple explicit examples of these would have been helpful, and the sooner the better. Mahlers chapter is written with much authority and enthusiasm, as witnessed by his short history of multitarget filtering (page 24, Chapter 14) which is rather fun to read and should be compared with Larry Stones version of history (page 22, Chapter 10). Two chapters by Julier and Uhlmann, on covariance intersection and nonlinear systems are particularly well done, being lucid, practical, and innovative; there is no doubt about how to use this new theory, as the MATLAB source code is included. The second chapter describes a novel algorithm for nonlinear filtering that has vastly superior performance compared with the extended Kalman filter (EKF) for certain applications. The EKF uses a simple linearization of the nonlinear equations to approximate the propagation of uncertainty in nonlinear transformations, whereas the new filter uses a more accurate approximation, based on sampling the probability density at carefully chosen points to

approximate an n-dimensional integral, similar to Gauss-Hermite quadrature. This is called the unscented filter for obvious reasons. Covariance intersection (CI) is a simple method of combining covariance matrices for fusion, when there is uncertainty about the statistical correlations involved. Roughly speaking, CI computes the largest one-sigma error ellipsoid that fits within the intersection of the two error ellipsoids being fused. This is a more conservative approach than using the standard Kalman filter equations, which assume perfect knowledge of the statistical correlations. However, when the two error ellipsoids are equal, the intersection is the same as the original ellipsoids, and, hence, using CI there is no apparent benefit from fusion; some engineers might view this as being much too conservative. One of the longest and most provocative chapters in this handbook is by Joseph Carl on Bayesian vs. Dempster-Shafer (D-S) decision algorithms, which have been the subject of heated debates over the last several decades. Carl says that many would argue that probability theory is not suitable for practical implementation on complex real-world problems, which is a very interesting assertion, but, Carl does not explain how D-S theory improves the situation. In particular, it is well-known and not subject to debate that given a decision problem with completely defined probability distribution functions (pdfs), that Bayesian decision rules are optimal. Therefore, the only way in which D-S could improve performance is in the case in which the pdfs are not completely specified. The problem of incomplete pdfs was the original motivation for Dempsters seminal work published in 1967 and 1968, and it has been the subject of extensive research since then, but without any practical results so far (see [6] and pages 57-61 in [7] for a survey of this research). The best work that I know of on this subject is Chapter 5 of Kharins recent book [8], which basically concludes, that for most sensible models of uncertainty in pdfs, the standard Bayesian decision rule, or a minor modification of it, is the most robust approach. There is no formula or theorem in Carls chapter that shows how much D-S improves performance relative to Bayesian methods, nor does Carl assert that D-S is better than Bayes, but rather in the detailed numerical example that is worked out, the D-S and Bayesian decisions are qualitatively the same. This agrees with other comparisons of D-S with Bayes reported in the

IEEE AESS Swfenis Muguzine. October 2001

15

literature, when such calculations are done correctly. There are, unfortunately, notorious cases of papers published by engineers in which the D-S advocate claimed superiority over Bayes, owing to the engineers faulty understanding of elementary probability theory. Moreover, it is well-known that Dempsters rule of combining evidence assumes that the data are statistically independent, whereas Bayesian methods do not require this assumption, with the result that for problems with statistically dependent data, the Bayesian performance can be vastly superior to D-S. This is a fundamental limitation of D-S compared with Bayes, but Carl does not mention it in his chapter, which is a curious lacuna in a handbook on data fusion such as this. A common misconception among engineers is that Bayesian methods require Gaussian pdfs, uniform prior distributions, and/or statistically independent data, whereas, in fact, Bayesian decision rules are optimal for any pdf of the data (Gaussian or non-Gaussian), any prior pdf (uniform or non-uniform) and data that are statistically dependent or independent, Three good references on D-S vs. Bayesian methods, which are both readable by normal engineers as well as mathematically correct are [6, 11 Chapter 9, 121. Carl mentions near the end of his chapter that D-S is not equipped with a clear-cut decision making rule, which many readers will find somewhat surprising, because D-S has been touted as a better method of decision making. For those wondering what it means to say that D-S has no decision rules, one should read chapters 13 to 17 in [9]. Moreover, in the Foreword to Shafers famous book [lo], Dempster very clearly says that: I differ from Shafer . . . I believe that Bayesian inference will always be a basic tool for practical everyday statistics, if only because questions must be answered. and decisions must be taken, So, Professor Dempster is apparently a Bayesian, and he would not advocate the use of D-S for decision making. About ten years ago, mathematicians gave up trying to make better decisions than Bayes; the recent academic papers on D-S say that there is more to life than making decisions, that Bayes wins at making decisions, and that maybe D-S is useful to represent uncertainty. However, hard-boiled engineers are only interested in making decisions; representing uncertainty is for poets, not engineers. Any reader interested in data fusion should study the authoritative and nicely written chapters by Professors Poore, BarShalom, Kirubarajan and Uhlmann (it only seems like Uhlmann wrote half of the chapters in this handbook), as well as by Julier, Mahler, and Stone, which provide a solid mathematical foundation for data fusion. However, this handbook would still benefit from a healthy dose of the three R s: Resolution, Registration and Real world applications. First, the issue of limited sensor resolution is briefly mentioned by Stone, but, otherwise, it is essentially ignored by the other authors. For typical multisensor data fusion applications, the problem of limited sensor resolution is often more important

than data association [l-31. Second, there is only one very short chapter on data registration (and this chapter is limited to images), but [:hiscrucial topic, often the Achilles heel of real world multismensor data fusion applications, deserves much more treatment in a handbook such as this [4-51. Third, very few chapters of this handbook describe significant real world applications or simulations of such applications. One yearns for a handbook of multisensor data fusion more like Skolnik!; Radar Handbook, in which nearly every chapter was written by an engineer with a firm grasp of theory as well as substantial real world experience, and which includes practical formulas, many useful plots, lessons learned from designing and testing real systems, but without a mClange of buzzwords, platitudes, and taxonomies. - Fred Daum, Raytheon

REFERENCES
[ I ] Koch, W. and Van Keuk, G . , 1997,
Multi,ple hypothesis track maintenance with possibly unresolved measurements, IEEE Transactions on Aerospace and Electronic Systems, 33,883-892. [2] Blair, W.D. and Brandt-Peace, M., July 1999, NNJPDA f o r tracking closely-spaced Rayleigh Targers wirh possibly merged measurements, Proceedings of SPIE Conference on Signal and Data Processing, Vol. 3809,396-408. [3] Daum, F.E. and Fitzgerald, R.J., April 1994, The importance of resolution in multiple target tracking. Proceedings of SPIE Conference on Signal and Data Processing, Vol. 2235,329-338.

[4] Dana, 1vI.P.. 1990, Regi:itration: a prerequisitef o r multiple sensor tracking, Chapter 5 in Multitarget-Multisensor Tracking, edited by Yaakov Bar-Shalom, Artech House. [5] Mori, S . and Chung, C.-Y., June 2000, Effects of unpaired objects and sensor biases on track-to-track association, Proceedings of MSS Sensor and Data Fusion, 137-151. [6] Walley, P., 1991, Staristical Reasoning with Imprecise Probabilities, Chapman and Hall. [7] RobusiBayesian Analysis, edited by D. lnsua and F. Ruggeri., Springer-Verlag, 2000. [SI Khariu, Y., 1996, Robustness in Staristical Pattern Recognition, Kluwer. [9] Advances in the Dempster-Shafer Theory of Evidence, edited by R.R. Yager, et al., John Wiley & Sons, 1994. [lo] Sharer, Q., 1976, A Mathematical Theory of Evidence, Princeton University Press. [ l 11 Pearl, J., 1988, Probabilistic Reasoning in Intelligent Systems, Morgan Kaufmann. [I21 Waserman, L., 1990, Belieffunctions and statistical inference, Canadian Joumal of Statistics.

16

IEEE AESS Swenu Muguzine. October 2001