Roberto Nerino

Vision-Based Pose Estimation for Robot-Mediated Hand Telerehabilitation

Vision-based Pose Estimation (VPE) represents a non-invasive solution to allow a smooth and natural interaction between a human user and a robotic system, without requiring complex calibration procedures. Moreover, VPE interfaces are gaining momentum as they are highly intuitive, such that they can be used from untrained personnel (e.g., a generic caregiver) even in delicate tasks as rehabilitation exercises. In this paper, we present a novel master–slave setup for hand telerehabilitation with an intuitive and simple interface for remote control of a wearable hand exoskeleton, named HX. While performing rehabilitative exercises, the master unit evaluates the 3D position of a human operator’s hand joints in real-time using only a RGB-D camera, and commands remotely the slave exoskeleton. Within the slave unit, the exoskeleton replicates hand movements and an external grip sensor records interaction forces, that are fed back to the operator-therapist, allowing a direct real-time assessment of the rehabilitative task. Experimental data collected with an operator and six volunteers are provided to show the feasibility of the proposed system and its performances. The results demonstrate that, leveraging on our system, the operator was able to directly control volunteers’ hands movements.

Image description using Gabor wavelets

A technique for the detection of characteristic points and local description of an image is presented. This technique is based on the use of Gabor wavelets and allows point-to- point correspondence between two images of the same scene taken from different points of view. In order to evaluate its usefulness, the technique is compared with the classic local correlation approach.

A novel approach to train random forests on GPU for computer vision applications using local features

The random forests (RF) classifier has recently gained momentum in the computer vision field, thanks to its successful application in human body tracking, hand pose estimation and object detection. In this article, we present a novel approach to train RF on a graphics processing unit (GPU) for computer vision applications where simple per-pixel features are computed. Besides leveraging the processing power of the GPU to accelerate the training, we reformulate the training problem to limit costly image transfers when it is not possible to store the entire data set in GPU memory. Furthermore, our implementation supports arbitrary image types and allows the user to specify custom features. We extensively compare our approach with the state of the art on publicly available data sets, and we obtain a reduction in training time of up to 18 times. Finally, we train our implementation on a large data set (around 100 K images), demonstrating that our approach is suitable for training RF on the vast data sets typically used in computer vision.

An Integrated Approach to the Well-Being of the Elderly People at Home

The paper presents the outline and the preliminary developments of NINFA (iNtelligent Integrated Network For Aged people), a project for the well-being of the elderly people at home. This architecture is based on a service platform suited for elder people called the Virtual Village Network, whose user interface allows to deliver different services at home, namely: user supervision, communication and interaction among users for social inclusion, exergame delivering, monitoring of the wellness status. After the discussion of some results of the investigation on the acceptability issues of the ICT technologies related to the project, the User Interface (UI) and the novel Human Computer Interface (HCI) have been developed. The HCI is particularly suited for elderly people and motor impaired patients because the interaction is managed only by finger/hand gestures and by vocal control through simple commands. A set of preliminary exergames developed for the user training and monitoring are presented. During the exergame execution, the user interface allows the real-time acquisition of a set of motor, linguistic and cognitive parameters related to the user performance. The analysis of the verbal production of each subject is used to observe its language evolution and to detect the onset of any cognitive deficit. The relationship between some parameters and the neurological/wellness status of the user is discussed.

A Self-Managed System for Automated Assessment of UPDRS Upper Limb Tasks in Parkinson’s Disease

A home-based, reliable, objective and automated assessment of motor performance of patients affected by Parkinson’s Disease (PD) is important in disease management, both to monitor therapy efficacy and to reduce costs and discomforts. In this context, we have developed a self-managed system for the automated assessment of the PD upper limb motor tasks as specified by the Unified Parkinson’s Disease Rating Scale (UPDRS). The system is built around a Human Computer Interface (HCI) based on an optical RGB-Depth device and a replicable software. The HCI accuracy and reliability of the hand tracking compares favorably against consumer hand tracking devices as verified by an optoelectronic system as reference. The interface allows gestural interactions with visual feedback, providing a system management suitable for motor impaired users. The system software characterizes hand movements by kinematic parameters of their trajectories. The correlation between selected parameters and clinical UPDRS scores of patient performance is used to assess new task instances by a machine learning approach based on supervised classifiers. The classifiers have been trained by an experimental campaign on cohorts of PD patients. Experimental results show that automated assessments of the system replicate clinical ones, demonstrating its effectiveness in home monitoring of PD.

Pose estimation of an active stereo system by principal moments of point features

The estimation of the pose variation (egomotion) of a vision system while looking at the same scene from two different poses is both of theoretical and practical interest in computer vision. In this work we deal with this problem by developing a method to estimate the relative motion of our stereo vision system with respect to an object surface in a static scene. The images of the object taken from different views are described by a set of well localized 2D features named key points with associated local descriptors. These descriptors are based on coefficients of a Gabor wavelet transform of the images, and on the 3D object co-ordinates of the corresponding points in the space. 3D point co- ordinates are evaluated by means of a fixation process performed by our active stereo vision system. A selection process of key points based on the similarity of their descriptors in the two images produces two sets with the same number of key points, every set having a high probability to be composed by the correspondent points in the other images. The centroids and the principal moment axes (roto-translation invariants) of the 3D point co- ordinates of the two sets are evaluated, and the relative motion between the two poses is recovered from this information, thus avoiding a direct point to point matching.