Classification of Streetsigns Using Gaussian Process Latent Variable Models

Since the rise of deep artificial neuronal nets, object detection and classification became an autonomous procedure, where both, feature extraction and feature processing (e.g.: classification) is done using an architecture based on artificial neurons. The shortcomings of deep neuronal nets are mainly based the black box models and the architecture of the networks, which cannot be estimated. Unknown behavior and over-fitting is still an unsolved problem. Thus, human-made parameters like the number of neurons or the definition of activation functions must be set. This work presents a non-parametric and non-linear approach for image processing using latent variable models. We used Gaussian process latent variable models for street sign feature extraction, where a latent representation is estimated without prior knowledge such as class label. Based on the latent representation, we visualizes the features and use state-of-the-art classifier for street sign classification. Our results proves, that our approach extracts useful features for classification. Our approach has still shortcomings, such as computational time, which are current areas of research.