Oliver Brdiczka

Automatic detection of interaction groups

This paper addresses the problem of detecting interaction groups in an intelligent environment. To understand human activity, we need to identify human actors and their interpersonal links. An interaction group can be seen as basic entity, within which individuals collaborate in order to achieve a common goal. In this regard, the dynamic change of interaction group configuration, i.e. the split and merge of interaction groups, can be seen as indicator of new activities. Our approach takes speech activity detection of individuals forming interaction groups as input. A classical HMM-based approach learning different HMM for the different group configurations did not produce promising results. We propose an approach for detecting interaction group configurations based on the assumption that conversational turn taking is synchronized inside groups. The proposed detector is based on one HMM constructed upon conversational hypotheses. The approach shows good results and thus confirms our conversational hypotheses.

Detecting individual activities from video in a smart home

This paper addresses the detection of activities of individuals in a smart home environment. Our system is based on a robust video tracker that creates and tracks targets using a wide-angle camera. The system uses target position, size and orientation as input for interpretation. Interpretation produces activity labels such as walking, standing, sitting, interacting with table, or sleeping for each target. Bayesian Classifier and Support Vector Machines (SVMs) are compared for learning and recognizing previously defined individual activities. These methods are evaluated on recorded data sets. A novel Hybrid Classifier is then proposed. This classifier combines generative Bayesian methods and discriminative SVMs. Bayesian methods are used to detect previously unseen activities, while the SVMs are shown to provide high discriminative power for recognizing examples of learned activity classes. The evaluation results of the Hybrid classifier for the recorded data sets show that the combination of generative and discriminative classification methods outperforms the individual methods when identifying unseen activities.

Learning situation models for providing context-aware services

In order to provide information and communication services without disrupting human activity, information services must implicitly conform to the current context of human activity. However, the variability of human environments and human preferences make it impossible to preprogram the appropriate behaviors for a context aware service. One approach to overcoming this obstacle is to have services adapt behavior to individual preferences though feedback from users. This article describes a method for learning situation models to drive context-aware services. With this approach an initial simplified situation model is adapted to accommodate user preferences by a supervised learning algorithm using feedback from users. To bootstrap this process, the initial situation model is acquired by applying an automatic segmentation process to sample observation of human activities. This model is subsequently adapted to different operating environments and human preferences through interaction with users, using a supervised learning algorithm.

Automatic Acquisition of Context Models and its Application to Video Surveillance

This paper addresses the problem of automatically acquiring context models from data. Context and human behavior are represented using a state model, called situation model. This model consists of different layers referring to entities, filters, roles, relations, situation and situation relationship. We propose a framework for the automatic acquisition of these different layers. In particular, this paper proposes a novel generic situation acquisition algorithm. The algorithm is also successfully applied to a video surveillance task and is evaluated by the public CAVIAR video database. The results are encouraging

Supervised learning of an abstract context model for an intelligent environment

This paper addresses the problem of supervised learning in intelligent environments. An intelligent environment perceives user activity and offers a number of services according to the perceived information about the user. An abstract context model in the form of a situation network is used to represent the intelligent environment, its occupants and their activities. The context model consists of situations, roles played by entities and relations between these entities. The objective is to adapt the system services, which are associated to the situations of the model, to the changing needs of the user. For this, a supervisor gives feedback by correcting system services that are found to be inappropriate to user needs. The situation network can be developed by exchanging the system service-situation association, by splitting the situation, or by learning new roles. The situation split is interpreted as a replacement of the former situation by sub-situations whose number and characteristics are determined using conceptual or decision tree algorithms. Different algorithms have been tested on a context model within the SmartOffice environment of the PRIMA research group. The decision tree algorithm (ID3) has been found to give the best results.

Context-aware environments: from specification to implementation

: This paper deals with the problem of implementing a context model for a smart environment. The problem has already been addressed several times using many different data- or problem-driven methods. In order to separate the modelling phase from implementation, we first represent the context model by a network of situations. Then, different implementations can be automatically generated from this context model depending on user needs and underlying perceptual components. Two different implementations are proposed in this paper: a deterministic one based on Petri nets and a probabilistic one based on hidden Markov models. Both implementations are illustrated and applied to real-world problems.

Detecting small group activities from multimodal observations

AbstractnThis article addresses the problem of detecting configurations and activities of small groups of people in an augmented environment. The proposed approach takes a continuous stream of observations coming from different sensors in the environment as input. The goal is to separate distinct distributions of these observations corresponding to distinct group configurations and activities. This article describes an unsupervised method based on the calculation of the Jeffrey divergence between histograms over observations. These histograms are generated from adjacent windows of variable size slid from the beginning to the end of a meeting recording. The peaks of the resulting Jeffrey divergence curves are detected using successive robust mean estimation. After a merging and filtering process, the retained peaks are used to select the best model, i.e. the best allocation of observation distributions for a meeting recording. These distinct distributions can be interpreted as distinct segments of group configuration and activity. To evaluate this approach, 5 small group meetings, one seminar and one cocktail party meeting have been recorded. The observations of the small groups meetings and the seminar were generated by a speech activity detector, while the observations of the cocktail party meeting were generated by both the speech activity detector and a visual tracking system. The authors measured the correspondence between detected segments and labeled group configurations and activities. The obtained results are promising, in particular as the method is completely unsupervised.n

Deterministic and probabilistic implementation of context

This paper addresses the problem of implementing an abstract context model. First, the abstract context model is represented by a network of situations. Two different implementations for the situation model are then proposed: a deterministic one based on Petri nets and a probabilistic one based on hidden Markov models. Both implementations are illustrated and applied to real-world problems

Unsupervised Segmentation of Meeting Configurations and Activities using Speech Activity Detection

This paper addresses the problem of segmenting small group meetings in order to detect different group configurations and activities in an intelligent environment. Our approach takes speech activity detection of individuals attending a meeting as input. The goal is to separate distinct distributions of speech activity observation corresponding to distinct group configurations and activities. We propose an unsupervised method based on the calculation of the Jeffrey divergence between histograms of speech activity observations. These histograms are generated from adjacent windows of variable size slid from the beginning to the end of a meeting recording. The peaks of the resulting Jeffrey divergence curves are detected using successive robust mean estimation. After a merging and filtering process, the retained peaks are used to select the best model, i.e. the best speech activity distribution allocation for a given meeting recording. These distinct distributions can be interpreted as distinct segments of group configuration and activity. To evaluate, we recorded 6 small group meetings. We measured the correspondence between detected segments and labeled group configurations and activities. The obtained results are promising, in particular as our method is completely unsupervised.

Learning context models for the recognition of scenarios

This paper addresses the problem of automatic learning of scenarios. A ubiquitous computing environment must have the ability to perceive its occupants and their activities in order to recognize a context and to provide appropriate services. A context (a scenario) can be modeled as a temporal sequence of situations. Hard coding contexts by hand is a complex task. Our goal is to learn these context models based on a set of videos showing actors playing predefined scenarios. Once these models are learned, we can use them to classify new scenarios. Hidden Markov Models (HMMs) are particularly well suited for problems with a strong temporal structure; they are easily adaptable to variability of input and robust to noise. But two problems need to be addressed: how many HMMs do we need for all possible scenarios and how many states for each HMM. We propose in this paper an approach based on an incremental algorithm addressing these two problems. Under the best conditions we obtained the minimal error rate of 1.96% (2 errors in 102 validation entries).