Object Classification using Probabilistic Graphical Model

Abstract

In the field of autonomous driving a lot of tasks which were solved by humans must be solved by algorithms instead. A common task is the classification of objects and the semantic understanding of a scene. This thesis tries to solve the classification task with an approach from the research field of Statistical Relational Learning (SRL). The first goal of the work is to analyze the usability of ontologies in the field of autonomous driving. The second goal of the work is to improve the classification results and runtime of a probabilistic approach from [Hen03] by considering the information of time. A concept of three consecutive steps was developed. To solve the task semantic knowledge which can be used to classify the objects has to be available for the system. Firstly this knowledge is modelled with an ontology. Secondly the different classifiers are developed and trained. These classifiers are graphical models from the field of SRL, more precisely Markov Logic Network (MLN). Thirdly the dfferent classifiers are compared and evaluated. The usability of ontologies could not be finally shown but the classification was improved, both in quality and runtime.