FBK home > INFORMATION TECHNOLOGY > Technologies of Vision > People > E.Ricci > research

Technologies of Vision

Current projects

Alpha Tracking Multiple People with Illumination Maps

Realtime multi-object tracking in an illumination-varying environment is a classical task in computer vision. Many approaches have been proposed in the literature but the problem is still far from being solved. In order to obtain robust tracking results, some methods simply discard the illumination-sensitive color information and employ other features that are considered invariant to illumination such as edges or textures. However, the main problem with such approaches is that in case of cluttered background, edges or textures are often not sufficient for reliably differentiating moving object contours from their background. Other approaches still rely on color information. However in order to handle illumination changes a common strategy is to adopt a color space different from RGB such as YUV or HSI in order to eliminate the intensity component. The shortcoming is that the feature discrimination capability is reduced, since only parts of the color channels are used. In the domain of tracking with particle filters, a scarcely investigated but promising method for dealing with varying illuminations conditions consists in a unified approach for jointly estimating the positions of the targets and their illumination conditions. The motivation behind this is that target localization strongly depends on object appearance and at the same time illumination conditions of a target are influenced by its position in the scene. Starting from this idea, in this project, we aim to develop a new algorithm for visual tracking of multiple people under non-homogenous and time-varying illumination conditions.

Learning People Trajectories Learning Pedestrian Trajectories -->

The automatic analysis of usual patterns is crucial for many video surveillance applications such as visual object tracking or anomaly detection. A typical example is represented by pedestrian trajectories. Usually, pedestrians tend to follow only a few common trajectories whilst other paths are very infrequent or never observed. In this project we aim to investigate novel approches for learning pedestrian trajectories.

Past projects

vfoa Estimating the Visual Focus of Attention of People in Realtime

The goal of this project was the study of the Visual Focus of Attention (VFoA) of a person ("where is he/she looking at?") in a meeting scenario. In the context of the analysis of non-verbal communication in meetings, my postdoc activity at Idiap Research Institute focused on the design and the implementation of a close to real-time head tracking and pose estimation system to be used for estimating the visual focus of attention of meeting participants. This project was developed in the framework of the AMI/AMIDA project, a EU project jointly lead by Idiap and the University of Edinburgh, which addresses the general issue of computer enhanced multi-modal interaction in the context of meetings.

vfoa Magic Moments for Structured Output Prediction

The project was about the study of discriminative methods for learning with structured data such as sequences or graphs. Learning in structured output refers to a topic which has become popular in machine learning and aims, to a large extent, to unify kernel methods and probabilistic graphical models into a single framework. Structured output predictions algorithms deal with problems involving structured data, moving out from the conventional and very limiting assumption that the data are expressed as vectors and that prediction amounts to assigning them to one of a few classes. Most approaches to structured output prediction rely on a hypothesis space of prediction functions that compute their output by maximizing a linear scoring function. We developed a new approach for structured output prediction problems together with a statistical analysis of its performance. The method relies on efficiently computing the first two moments of the scoring function over the output space, and using them to create convex objective functions for training. We tested the proposed approach in several tasks such as sequence labeling in natural language processing, DNA sequence alignment and RNA folding in bioinformatics.

vfoa Automatic Analysis of Retinal Fundus Images

The project involved mainly the study of novel approaches for the automatic processing of retinal fundus images to aid the diagnosis and the treatment of eye diseases. We developed new algorithms mainly based on kernel methods to extract relevant features such as blood vessels and pathological elements in retinal images.