Search

The focus of this work is emotion recognition in the wild based on a multitude of different audio, visual and meta features. For this, a method is proposed to optimize multi-modal fusion architectures based on evolutionary computing. Extensive uni- and multi-modal experiments show the discriminative power of each computed feature set and fusion architecture. Furthermore, we summarize the EmotiW 2013/2014 challenges and review the conclusions that have been drawn and compare our results with the state-of-the-art on this dataset.

The recognition of patterns in real-time scenarios has become an important trend in the field of multi-modal user interfaces in human computer interaction. Cognitive technical systems aim to improve the human computer interaction by means of recognizing the situative context, e.g. by activity recognition (Ahad et al. in IEEE, 1896–1901, 2008), or by estimating the affective state (Zeng et al., IEEE Trans Pattern Anal Mach Intell 31(1):39–58, 2009) of the human dialogue partner. Classifier systems developed for such applications must operate on multiple modalities and must integrate the available decisions over large time periods. We address this topic by introducing the Markov fusion network (MFN) which is a novel classifier combination approach, for the integration of multi-class and multi-modal decisions continuously over time. The MFN combines results while meeting real-time requirements, weighting decisions of the modalities dynamically, and dealing with sensor failures. The proposed MFN has been evaluated in two empirical studies: the recognition of objects involved in human activities, and the recognition of emotions where we successfully demonstrate its outstanding performance. Furthermore, we show how the MFN can be applied in a variety of different architectures and the several options to configure the model in order to meet the demands of a distinct problem.

The individual nature of physiological measurements of human affective states makes it very difficult to transfer statistical classifiers from one subject to another. In this work, we propose an approach to incorporate unlabeled data into a supervised classifier training in order to conduct an emotion classification. The key idea of the method is to conduct a density estimation of all available data (labeled and unlabeled) to create a new encoding of the problem. Based on this a supervised classifier is constructed. Further, numerical evaluations on the EmoRec II corpus are given, examining to what extent additional data can improve classification and which parameters of the density estimation are optimal.

The analysis of affective or communicational states in human-human and human-computer interaction (HCI) using automatic machine analysis and learning ap- proaches often suffers from the simplicity of the approaches or that very ambitious steps are often tried to be taken at once. In this paper, we propose a generic framework that overcomes many difficulties associated with real world user behavior analysis (i.e. uncertainty about the ground truth of the current state, subject independence, dynamic realtime analysis of multimodal information, and the processing of incomplete or erroneous inputs, e.g. after sensor failure or lack of input). We motivate the approach, that is based on the analysis and spotting of behavioral cues that are regarded as basic building blocks forming user state specific behavior, with the help of related work and the analysis of a large HCI corpus. For this corpus paralinguistic and nonverbal behav- ior could be significantly associated with user states. Some of our previous work on the detection and classification of behavioral cues is presented and a layered architecture based on hidden Markov models is introduced. We believe that this step by step approach towards the understanding of human behavior underlined by encouraging preliminary results out- lines a principled approach towards the development and evaluation of computational mechanisms for the analysis of multimodal social signals.