Performance evaluation of fusion rules for multitarget tracking in clutter based on generalized data association

Abstract

In this paper, we present and compare different fusion rules which can be used for Generalized Data Association (GDA) for multitarget tracking (MIT) in clutter. Most of tracking methods including Target Identification (ID) or attribute information are based on classical tracking algorithms as PDAF, JPDAF, MHT, IMM, etc and either on the Bayesian estimation and prediction of target ID, or on fusion of target class belief assignments through the Demspter-Shafer Theory (DST) and Dempster's rule ofcombination. In this paper we pursue ourprevious works on the development of a new GDA-M7T based on Dezert-Smarandache Theory (DSmT) but compare also it with standard fusion rules (Demspter's, Dubois & Prade's, Yager's) and with a new fusion Proportional Conflict Redistribution (PCR) rule in order to assess the efficiency of all these differentfusion rules for this GDA-MTT in highly conflicting situation. This evaluation is based on a Monte Carlo simulation for a difficult maneuvering MI1 problem in clutter similar to the example recently proposed by Bar-Shalom, Kirubarajan and Gokberk.

Overview of Dempster-Shafer and belief function tracking methods

“…Beginning in 2005, efforts were made to extend traditional DS tracking methods [39] to that of advanced techniques using the proportional conflict redistribution rule (PCR5) [40]. Other tracking methods included multisensor [41], activity analysis [42], and out-of-sequence methods [43].…”

On the Quality of Optimal Assignment for Data Association

“…Applying the pignistic transformation 6 [24], we get finally BetP (X) = m(X)+ 1 2 • m(X ∪ ¬X) and BetP (¬X) = 1 2 • m(X ∪ ¬X). Therefore, we choose the quality indicator as q k2 II (i, j) = BetP (X).…”

“…Fortunately, efficient algorithms have been developed in operational research and tracking communities for formalizing and solving these optimal assignments problems. Several approaches based on different models can be used to establish rewards matrix, either based on the probabilistic framework [1,3], or on the belief function (BF) framework [4][5][6][7]. In this paper, we do not focus on the construction of the rewards matrix 1 , and our purpose is to provide a method to evaluate the quality (interpreted as a confidence score) of each association (pairing) provided in the optimal solution based on its consistency (stability) with respect to all the second best solutions.…”

The Impact of the Quality Assessment of Optimal Assignment for Data Association in a Multitarget Tracking Context

“…Some of them establish a reward matrix based on Kinematic only Data Association (KDA) and on a probabilistic framework [3,4]. Some of them rely on Belief Functions (BF) [5][6][7][8][9] and motivate the incorporation of the advanced concepts for Generalized Data Association (GDA) [6][7][8], allowing the introduction of a target attribute (target type, radar cross section, etc.) into the association logic, in order to improve the track maintenance performance in complicated situations (closely spaced/crossing targets), when kinematics data are insufficient for coherent decision making.…”