Robust and Distributed Cluster Enumeration and Object Labeling

• Main Author: Teklehaymanot, Freweyni Kidane
• Format: Others
• Language: en
• Published: 2019
• Online Access: https://tuprints.ulb.tu-darmstadt.de/8539/1/2019-03-04_Teklehaymanot_Freweyni_Kidane.pdf; Teklehaymanot, Freweyni Kidane <http://tuprints.ulb.tu-darmstadt.de/view/person/Teklehaymanot=3AFreweyni_Kidane=3A=3A.html> (2019): Robust and Distributed Cluster Enumeration and Object Labeling.Darmstadt, Technische Universität, [Ph.D. Thesis]

This dissertation contributes to the area of cluster analysis by providing principled methods to determine the number of data clusters and cluster memberships, even in the presence of outliers. The main theoretical contributions are summarized in two theorems on Bayesian cluster enumeration based on modeling the data as a family of Gaussian and t distributions. Real-world applicability is demonstrated by considering advanced signal processing applications, such as distributed camera networks and radar-based person identification. In particular, a new cluster enumeration criterion, which is applicable to a broad class of data distributions, is derived by utilizing Bayes' theorem and asymptotic approximations. This serves as a starting point when deriving cluster enumeration criteria for specific data distributions. Along this line, a Bayesian cluster enumeration criterion is derived by modeling the data as a family of multivariate Gaussian distributions. In real-world applications, the observed data is often subject to heavy tailed noise and outliers which obscure the true underlying structure of the data. Consequently, estimating the number of data clusters becomes challenging. To this end, a robust cluster enumeration criterion is derived by modeling the data as a family of multivariate t distributions. The family of t distributions is flexible by variation of its degree of freedom parameter (ν) and it contains, as special cases, the heavy tailed Cauchy for ν = 1, and the Gaussian distribution for ν → ∞. Given that ν is sufficiently small, the robust criterion accounts for outliers by giving them less weight in the objective function. A further contribution of this dissertation lies in refining the penalty terms of both the robust and Gaussian criterion for the finite sample regime. The derived cluster enumeration criteria require a clustering algorithm that partitions the data according to the number of clusters specified by each candidate model and provides an estimate of cluster parameters. Hence, a model-based unsupervised learning method is applied to partition the data prior to the calculation of an enumeration criterion, resulting in a two-step algorithm. The proposed algorithm provides a unified framework for the estimation of the number of clusters and cluster memberships. The developed algorithms are applied to two advanced signal processing use cases. Specifically, the cluster enumeration criteria are extended to a distributed sensor network setting by proposing two distributed and adaptive Bayesian cluster enumeration algorithms. The proposed algorithms are applied to a camera network use case, where the task is to estimate the number of pedestrians based on streaming-in data collected by multiple cameras filming a non-stationary scene from different viewpoints. A further research focus of this dissertation is the cluster membership assignment of individual data points and their associated cluster labels given that the number of clusters is either prespecified by the user or estimated by one of the methods described earlier. Solving this task is required in a broad range of applications, such as distributed sensor networks and radar-based person identification. For this purpose, an adaptive joint object labeling and tracking algorithm is proposed and applied to a real data use case of pedestrian labeling in a calibration-free multi-object multi-camera setup with low video resolution and frequent object occlusions. The proposed algorithm is well suited for ad hoc networks, as it requires neither registration of camera views nor a fusion center. Finally, a joint cluster enumeration and labeling algorithm is proposed to deal with the combined problem of estimating the number of clusters and cluster memberships at the same time. The proposed algorithm is applied to person labeling in a real data application of radar-based person identification without prior information on the number of individuals. It achieves comparable performance to a supervised approach that requires knowledge of the number of persons and a considerable amount of training data with known cluster labels. The proposed unsupervised method is advantageous in the considered application of smart assisted living, as it extracts the missing information from the data. Based on these examples, and, also considering the comparably low computational cost, we conjuncture that the proposed methods provide a useful set of robust cluster analysis tools for data science with many potential application areas, not only in the area of engineering.