Decision-level data fusion

Xiaoxun Zhu

doi:10.17918/00007883

Research in data fusion is motivated by the potential for significant enhancement in multisensor and multi-algorithm decision making. Developed initially for radar applications, data fusion is now studied for general-purpose signal processing, diversity in communication, and high performance computing. In part I of this dissertation, we provide an overview of sensor fusion, focusing on a specific application-mobile robot navigation. We review the main techniques for sensor fusion in robot navigation, emphasizing algorithms for robot self location. The review provides an arsenal of tools for addressing this (rather ill-posed) problem in machine intelligence, including Kalman filtering, rule-based techniques, behavior-based algorithms, and approaches that borrow from information theory, Dempster-Shafer reasoning, fuzzy logic and neural networks. In part II, we investigate the parallel decision fusion architecture whose necessity was established in part I. This architecture comprises N local detectors (LDs) and a decision fusion center (DFC). The local detectors observe a common volume of surveillance. The i^[th] detector uses its observation vector, z_i [belongs to] R^M, in a local decision rule f_i : R^M [right arrow] {-1,1} : z_i [right arrow] u_i, to determine whether to accept the null hypothesis H₀ (u_i = -1), or the alternative hypothesis H₁ (u_i = 1). The local observations are assumed to be statistically independent, conditioned on the hypothesis. The local decision vector {u_i}^N_[i=1] is then transmitted to the DFC which uses a global decision rule f₀ : {-1,1}^N [right arrow] {-1,1} : {u_i}^N_[i=1] [right arrow] u₀ to generate the global decision, u₀. We extend the existing optimal design in two directions: (1) Adaptive decision fusion: when a priori probabilities and sensor characteristics are unknown, we devise an algorithm that learns the decision rule at the DFC on line. (2) M-ary hypothesis testing: when the local detectors are restricted to transmit only one bit (1 or -1), we provide the condition under which the probability of error at the DFC converges to zero asymptotically. We point to several future research needs, among them-adaptive algorithms, qualitative and quantitative issues, generalization of decision fusion systems, and hierarchical architectures.

Decision-level data fusion

Files and links (1)

Abstract

Metrics

Details

Decision-level data fusion

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media