Data fusion is the methodology of efficiently combining the relevant information from different sources. The goal is to achieve estimates and inferences with better confidence than those achievable by relying on a single source. Initial data fusion applications were predominantly in defense: target tracking, threat assessment and land mine detection. Nowadays, data fusion is applied to robotics (e.g., environment identification for navigation), medicine (e.g., medical diagnosis), geoscience (e.g., data integration from different sources) and industrial engineering (e.g., fault detection). This thesis focuses on data fusion for distributed multisensor tracking systems. In these systems, each sensor can provide the information as measurements or local estimates, i.e., tracks. The purpose of this thesis is to advance the research in the fusion of local estimates for multisensor multitarget tracking systems, namely, track fusion. This study also proposes new methods for track-to-track association, which is an implicit subproblem of track fusion. The first contribution is for the case where local sensors perform tracking using particle filters (Monte Carlo based methods). A method of associating tracks estimated through labeled particle clouds is developed and demonstrated with subsequent fusion. The cloud-to-cloud association cost is devised together with computation methods for the general and specialized cases. The cost introduced is proved to converge (with increasing clouds cardinality) toward the corresponding distance between the underlying distributions. In order to simulate the method introduced, a particle filter labeled at particle level was developed, based on the Probability Hypothesis Density (PHD) particle filter. The second contribution is for the case where local sensors produce tracks using Kalman filter-type estimators, in the form of track state estimate and track state covariance matrix. For this case the association and fusion is improved in both terms of accuracy and identity, by introducing at each fusion time the prior information (both estimate and identity) from the previous fusion time. The third contribution is for the case where local sensors produce track estimates under the form of MHT, therefore where each local sensor produces several hypotheses of estimates. A method to use the information from other sensors in propagating each sensor's internal hypotheses over time is developed. A practical fusion method for real world local tracking sensors, i.e., asynchronous and with incomplete information available, is also developed in this thesis. / Thesis / Doctor of Philosophy (PhD)

In this thesis, the Recursive-Random Sample Consensus (R-RANSAC) algorithm is applied to a vision-based, cooperative target tracking system. Unlike previous applications, which focused on a single camera platform tracking targets in the image frame, this work uses multiple camera platforms to track targets in the inertial or world frame. The process of tracking targets in the inertial frame is commonly referred to as geolocation.In practical applications sensor biases cause the geolocated target estimates to be biased from truth. The method for cooperative estimation developed in this thesis first estimates the relative rotational and translational biases that exist between tracks from different vehicles. It then accounts for the biases and performs the track-to-track association, which determines if the tracks originate from the same target. The track-to-track association is based on a sliding window approach that accounts for the correlation between tracks sharing common process noise and the correlation in time between individual estimation errors, yielding a chi-squared distribution. Typically, accounting for the correlation in time requires the inversion of a Nnx x Nnx covariance matrix, where N is the length of the window and nx is the number of states. Note that this inversion must occur every time the track-to-track association is to be performed. However, it is shown that by making a steady-state assumption, the inverse has a simple closed-form solution, requiring the inversion of only two nx x nx matrices, and can be calculated offline. Distributed data fusion is performed on tracks where the hypothesis test is satisfied. The proposed method is demonstrated on data collected from an actual vision-based tracking system.A novel method is also developed to cooperatively estimate the location and size of occlusions. This capability is important for future target tracking research involving optimized path planning/gimbal pointing, where a geographical map is unavailable. The method is demonstrated in simulation.

Covariance Matrix in target tracking algorithms has a critical role at multi- sensor track fusion systems. This matrix reveals the uncertainty of state es- timates that are obtained from diferent sensors. So, many subproblems of track fusion usually utilize this matrix to get more accurate results. That is why this matrix should be interchanged between the nodes of the multi-sensor tracking system. This thesis mainly deals with analysis of approximations of the covariance matrix that can best represent this matrix in order to efectively transmit this matrix to the demanding site. Kullback-Leibler (KL) Distance is exploited to derive some of the representations for Gaussian case. Also com- parison of these representations is another objective of this work and this is based on the fusion performance of the representations and the performance is measured for a system of a 2-radar track fusion system.

Multiple target tracking (MTT) is the process of localizing targets in an environment using sensors that perceive the environment. MTT has many applications such as wildlife monitoring, air traffic monitoring, and surveillance. These applications motivate further research in the different challenging aspects of MTT. One of these challenges that we will focus on in this dissertation is constructing a high fidelity target model. A common approach to target modeling is to use linear models or other simplified models that do not properly describe the target's pose (position and orientation), motion, and uncertainty. These simplified models are typically used because they are easy to implement and computationally efficient. A more accurate approach that improves tracking performance is to define the target model using a geometric representation of the target's natural configuration manifold. In essence, this geometric approach seeks to define a target model that can express every pose and motion of the target while preserving geometric properties such as distances and angles. We restrict our discussion of MTT to objects that move in physical space and can be modeled as a rigid body. This restriction allows us to construct generic geometric target models defined on Lie groups. Since not every Lie group has additional structure that permits vector space arithmetic like Euclidean space, many components of MTT such as data association, track initialization, track propagation and updating, track association and fusing, etc, must be adapted to work with Lie groups. The main contribution of this dissertation is the presentation of a novel MTT algorithm that implements the different MTT components to work with target models defined on Lie groups. We call this new algorithm, Geometric Multiple Target Tracking (G-MTT). This dissertation also serves as a guide on how other MTT algorithms can be modified to work with geometric target models. As part of the presentation there are various experimental results that strengthen the argument that a geometric approach to target modeling improves tracking performance.