Andreas Stergiou

Impact of Face Registration Errors on Recognition

Face recognition systems detect faces in moving or still images and then recognize them. However, face detection is not an error-free process, especially when designed for real-time systems. Thus the face recognition algorithms have to operate on faces that are not ideally framed. In this paper we analyze quantitatively the impact of face detection errors on six different face recognition algorithms. Hence, we propose a matching of face recognition algorithms with face detector performance, which can be used for a system based on the expected performance of the face detector.

A decision fusion system across time and classifiers for audio-visual person identification

In this paper the person identification system developed at Athens Information Technology is presented. It comprises of an audio-only (speech), a video-only (face) and an audiovisual fusion subsystem. Audio recognition is based on the Gaussian Mixture modeling of the principal components of the Mel-Frequency Cepstral Coefficients of speech. Video recognition is based on linear subspace projection methods and temporal fusion using weighted voting on the results. Audiovisual fusion is done by fusing the unimodal identities into the multimodal one, using a suitable confidence metric for the results of the unimodal classifiers.

The AIT Multimodal Person Identification System for CLEAR 2007

This paper presents the person identification system developed at Athens Information Technology and its performance in the CLEAR 2007 evaluations. The system operates on the audiovisual information (speech and faces) collected over the duration of gallery and probe videos. It comprises of an audio-only (speech), a video-only (face) and an audiovisual fusion subsystem. Audio recognition is based on the Gaussian Mixture modeling of the principal components of composite feature vectors, consisting of Mel-Frequency Cepstral Coefficients and Perceptual Linear Prediction coefficients of speech. Video recognition is based on combining three different classification algorithms: Principal Components Analysis with a modified Mahalanobis distance, sub-class Linear Discriminant Analysis (featuring automatic sub-class generation) with cosine distance and Bayesian classifier based on Gaussian modeling of intrapersonal differences. A nearest neighbor classification rule is applied. A decision fusion scheme across time and classifiers returns the video identity. The audiovisual subsystem fuses the unimodal identities into the multimodal one, using a suitable confidence metric.

2D and 3D face localization for complex scenes

In this paper, we address face tracking of multiple people in complex 3D scenes, using multiple calibrated and synchronized far-field recordings. We localize faces in every camera view and associate them across the different views. To cope with the complexity of 2D face localization introduced by the multitude of people and unconstrained face poses, a combination of stochastic and deterministic trackers, detectors and a Gaussian mixture model for face validation are utilized. Then faces of the same person seen from the different cameras are associated by first finding all possible associations and then choosing the best option by means of a 3D stochastic tracker. The performance of the proposed system is evaluated and is found enhanced compared to existing systems.

The AIT 2D Face Detection and Tracking System for CLEAR 2007

This paper describes the AIT system for 2D face tracking and the results obtained in the CLEAR 2007 evaluations. The system is based on the complementary operation of a set of face detectors and a deterministic tracker based on color. To minimize false positives, the system is applied on the body regions provided by a stochastic body tracker, and utilizes a detection validation scheme based on color and texture modeling of the faces.

The 2006 athens information technology speech activity detection and speaker diarization systems

This paper describes the Speech Activity Detection (SAD) and Speaker Diarization (SPKR) systems that were developed by the Athens Information Technology in the scope of the NIST RT-06S evaluations. The SAD system performs classification of recorded frames into speech and non-speech, using Linear Discriminant Analysis (LDA), while the SPKR one initially segments recordings into speech intervals based on the Bayesian Information Criterion (BIC), and then applies a two-step clustering strategy to group segments from the same speaker together. Following a discussion of the intrinsics of the two systems, we report and comment on our results on the RT-06S corpus [20].

Visual measurement cues for face tracking

Particle filters allow for visual trackers with nonlinear measurements. In this paper we consider three different non-linear visual measurement cues, based on object detection, foreground segmentation and colour matching. Novel ways to obtain robust measurement likelihoods under a unified representation scheme are discussed, followed by a likelihood combination scheme for fusion. The resulting single and multi-cue particle filter trackers are compared in the scope of face tracking.

The AIT Outdoor Tracker for Vehicles and Pedestrians in CLEAR2007

This paper presents the tracking system from Athens Information Technology that participated to the pedestrian and vehicle surveillance task of the CLEAR 2007 evaluations and the obtained results. The system is based on the CLEAR 2006 one, with some important modifications that are detailed. Since the test data in CLEAR 2006 and 2007 are the same, it is easy to quantify the obtained performance gain from the older system to the proposed one.