Technologies to automatically recognize stress, are extremely important to prevent chronic psychological stress and the pathophysiological risks associated to it. The introduction of comfortable and wearable biosensors have created new opportunities to measure stress in real-life environments, but there is often great variability in how people experience stress and how they express it physiologically. In this project, we explore the use of pattern recognition to predict stress levels of call center employees.

MIT Mood Meter is designed to assess and display the overall mood of the MIT community, by placing cameras at four different prime spots on the MIT campus. The cameras are equipped with affect-sensing software that counts number of people and whether they are smiling or not. This project is intended to raise awareness of how our own smiles can positively affect the surrounding environment, and to assess how congenial MIT is as a community. The dynamic, real-time information may lead to answers to questions such as: “Are people from one department happier than others?”, “Do midterms lower the mood?”, or “Does warmer weather lead to happiness?”

This project will define and explore a new research area we call Computational Behavior Science. More specifically, we hope to: (1) Enable widespread screening of autism by allowing non-experts to easily collect high-quality behavioral data and perform an initial assessment of risk status; (2) Improve behavioral therapy through increased availability and improved quality, by making it easier to track the progress of an intervention and follow guidelines for maximizing learning progress; and (3) Enable longitudinal analysis of a child's development based on quantitative behavioral data, using new tools for visualization.

Understanding if a movement is being executed correctly is important to a number of manipulation tasks. We present a preliminary study using inertial measurement units (IMUs) to evaluate the correct execution of squash swings, as well as discriminating them from similar racquet sport swings such as tennis and badminton. Additionally, we show some preliminary results that suggest that it is possible to detect at what point in time the movement was done incorrectly and why.

This project is funded by The Defense Advanced Research Projects Agency (DARPA) and seeks to create a "smart" computer that captures and evaluates ways in which expert users perform certain tasks. The computer then uses this knowledge in order to increase the productivity of non-expert users performing the same tasks.

The aim of this project is to develop an expert system capable of recognizing gunshots in audio files. We explore different types of audio features (e.g. Fast Fourier Transform, RASTA PLP) that are capable of characterizing the acoustic signatures of gunshots such as N-wave. Moreover, we compare the performance of Hidden Markov Models and Support Vector Machines.

Standard ensemble methods do not explicitly consider the performance of experts to be variable across the problem domain. In contrast, this work proposes Specialized Weighted Majority. Specialists for each area of the feature space are created by augmenting experts with a classifier that chooses on which samples to vote.

The hot flash is a sensation ranging from warmth to intense heat around the face, head, and neck. Populations experiencing hot flashes include male and female cancer patients, and most notably, women transitioning through menopause. Hot flashes are associated with impairments in quality of life and increased treatment seeking, and include contributions of both psychological and physiological factors in their experience and reporting. The goal of this project is to research temporal classifiers to predict hot flashes using several physiological measures (e.g. skin conductance, skin temperature, heat flux) and to understand which of the physiological factors are the most informative.

Complex knee injuries are common, often resulting from multiple forces (e.g. rotational, varus-valgus loading, anterior/posterior displacement). Identification of the specific injury pattern of the ACL (Anterior Cruciate Ligament) and other knee structures using non-invasive methods may improve pre-operative planning and guide treatment, reducing costs and facilitating high quality patient care. The main goal of this project is to present a classification system based on a set of non-invasive measures and state-of-the-art Machine Learning techniques to preempt the exact ACL rupture pattern. Standard classification techniques as Support Vector Machines (SVM) and Genetics Algorithms are combined for this purpose.

The purpose of this work is to explore the use of Conditional Random Fields (CRFs) in the automatic recognition of the basic units of facial expression (aka. Action Units). Facial expression recognition is a complex problem aggravated by variability between the subjects and variability in the dynamics of expression. The discriminative nature of CRFs and their ability to capture temporal dependencies makes them well suited to tackle these two difficulties. To aid the recognition of subtle expressions, a combination of shape and appearance features is used.

This work intends to develop a new kernel-based algorithm for horse recognition. The intuition behind this approach is that each horse has a set of distinctive features which may be different from other horses such as facial markings or color. With a unique structured sparse regularization, the proposed algorithm can be formulated as a non-smooth convex optimization problem.

Identifying facial features (e.g. eye corners) on images is an important task for many computer vision applications like face recognition, facial expression recognition, tracking, and 3D face modeling from 2D images. Some of the current methods to detect facial keypoints are based on sophisticated classifiers. This project proposes a purely data-driven approach based on template matching. The novelty of this method is finding a final template which will be a linear combination of other templates.

This project aims to perform transparent object segmentation on cluttered environments from a sequence of images. The main idea is that standard algorithms such as MeanShift do not perform well with transparent objects due to their properties. Because of that, we use standard algorithms to detect possible outliers and find the transparent object.

This project aims to develop a customized solution for managing the human resources of the 'Research Group in Intelligent Systems' from Ramon Llull University. With the objective of facilitating access to the application by all of its members, this project proposes a client-server implementation using LAMP (Linux + Apache + MySql + PhP). This project became my B.S thesis and earned an A with honors.