The Discriminative Generalized Hough Transform for Localization of Highly Variable Objects and its Application for Surveillance Recordings
Georg Hahmann · MACAU: Open Access Repository of Kiel University (University Library Kiel) · 2021
This work is about the localization of arbitrary objects in 2D images in general and the localization of persons in video surveillance recordings in particular. More precisely, it is about localizing specific landmarks. Thereby the possibilities and limitations of localization approaches based on the Generalized Hough Transform (GHT), especially of the Discriminative Generalized Hough Transform (DGHT) will be evaluated. GHT-based approaches determine the number of matching model and feature points and the most likely target point position is given by the highest number of matching model and feature points. Additionally, the DGHT comprises a statistical learning approach to generate optimal DGHT-models achieving good results on medical images. This work will show that the DGHT is not restricted to medical tasks but has issues with large target object variabilities, which are frequent in video surveillance tasks. As all GHT-based approaches also the DGHT only considers the number of matching model-feature-point-combinations, which means that all model points are treated independently. This work will show that model points are not independent of each other and considering them independently will result in high error rates. This drawback is analyzed and a universal solution, which is not only applicable for the DGHT but all GHT-based approaches, is presented. This solution is based on an additional classifier that takes the whole set of matching model-feature-point-combinations into account to estimate a confidence score. On all tested databases, this approach could reduce the error rates drastically by up to 94.9%. Furthermore, this work presents a general approach for combining multiple GHT-models into a deeper model. This can be used to combine the localization results of different object landmarks such as mouth, nose, and eyes. Similar to Convolutional Neural Networks (CNNs) this will split the target object variability into multiple and smaller variabilities. A comparison of GHT-based approaches with CNNs and a description of the advantages, disadvantages, and potential application of both approaches will conclude this work.