Random Finite Set Markov Chain Monte Carlo Predetection Fusion Ramona Georgescu Electrical and Computer Engineering Department University of Connecticut Storrs, CT Email: ramona@engr.uconn.edu Peter Willett Electrical and Computer Engineering Department University of Connecticut Storrs, CT Email: willett@engr.uconn.edu Abstract—Predetection fusion is an efficient (and, depending on what underlies it, indispensable) way to process high volume data from large networks of low quality sensors and thus, an aid to multisensor multitarget tracking. In previous work we derived both the GLRT (presumably “optimal”) technique and a more practicable contact-sifting variant. Unfortunately, the gaps between the two in terms of computation time and performance are not inconsiderable. Hence in this paper we propose a new approach based on random finite sets (RFS) and implemented by Monte Carlo (MCMC) simulation. We trust that it is found interesting; but even if not, we show that it offers improved results, in the sense of RMSE and number of declared targets. Keywords: Predetection Fusion, Markov Chain Monte Carlo, Random Finite Sets, Sensor Networks, Tracking. I. I NTRODUCTION The optimal sensor decision rule in the case of multiple sensor systems and known target location is known to be a likelihood ratio test [1]. This approach, however, is not applicable to many practical scenarios, such as sonar, in which the location of the target is not known and hence the alternative hypothesis becomes composite. We propose a practical implementation, Random Finite Set Markov Chain Monte Carlo (RFS MCMC) predetection fusion. Additional motivation comes from the fact that in recent years, interest has shifted towards deploying a large sensor network that consists of many but cheap, low quality sensors. Data fusion in large sensor networks is expected to provide better target tracking capability in terms of increased area coverage, expanded geometric diversity, increased target hold, robustness to sensor loss and jamming, improved localization and gains in probability of detection [2]. A possible drawback is an increased false alarm rate after the fusion step. A multistatic sonar system, which consists of multiple sonar sources and receivers distributed over the surveillance area, is one such sensor network. Transmissions from one or more sources may be processed by one or more receivers to produce a large number of sonar echo contacts [2]. The realistic multistatic sonar Metron [3] dataset provides a motivating example. Figure 1 shows the setup of this Metron dataset. There are 25 stationary sensors, all of them receivers with the exception of four sensors which are colocated source/receiver units; four targets display rectangular trajectories. The probability of detection is poor, on average P D =0.12 per sensor per scan. The high difficulty of this dataset is due to the extremely large number of contacts per scan and the low quality of the measurements. Figure 2 shows the first scan of data plotted in Cartesian coordinates: out of 890 contacts only 15 originate from a target, a major challenge to any tracking paradigm. Predetection fusion responds to this demanding data fusion problem by taking advantage of the benefits in both batch- and scan-based processing. RFS MCMC predetection fusion does so by blending multisensor measurements into a considerably smaller set of measurements to serve as input to a tracker, considerably reducing the number of false alarms while pre- serving most valid target detections. The technique is therefore beneficial to algorithms such as the Cardinalized Probability Hypothesis Density (CPHD) tracker [4] which is of O(nm 3 ) complexity, where n is the number of targets and m is the number of measurements. II. PREVIOUS WORK The optimal technique to solve the data fusion problem, the multihypothesis Generalized Likelihood Ratio Test (GLRT), has been proposed by Guerriero et al [5]. In previous work, we have introduced 2D predetection fusion, a practical approach that is many orders of magnitude faster than the GLRT. Here we present a short description of each method, and thereafter compare their performance against that of RFS MCMC predetection fusion. A. GLRT The natural way to tackle the problem of data fusion in large sensor networks is the GLRT approach, in which, for each hypothesized target, the location estimate which maximizes the likelihood function is found and the hypothesis with the largest likelihood is selected. Thus, the likelihood function is