Publications Lists     [ all BibTeX ]  

Hand motion tracking traditionally re- quires highly complex and expensive systems in terms of energy and computational demands. A low-power, low-cost system could lead to a revolution in this field as it would not require complex hardware or additional equipment. The present paper exploits the Multiple Point Tracking algorithm developed at the Tyndall National Institute as the basic algorithm to perform a series of gesture recognition tasks. The hardware relies upon the combination of a stereoscopic vision of two novel Lensless Smart Sensors (LSS) combined with IR filters and five hand-held LEDs to track. Track- ing common gestures generates a six-gestures dataset, which is then employed to train three Machine Learning models: k-Nearest Neighbors, Support Vector Machine and Random Forest. An offline analysis highlights how different LEDs’ positions on the hand affect the clas- sification accuracy. The comparison shows how the Random Forest outperforms the other two models with a classification accuracy of 90-91%. 

With the aim of monitoring human activities (in critical tasks as well as in leisure and sport activities), wearable devices provide enhanced usability and seamless human experience with respect to other portable devices (e.g. smartphones). At the same time, though, wearable devices are more resource-constrained in terms of computational capability and memory, which calls for the design of algorithmic solutions that explicitly take into account these issues. In this paper, a symbolic approach for activity recognition with wearable devices is presented: the Symbolic Aggregate approXimation technique is here extended to multi-dimensional time series, in order to capture the mutual information of different dimensions. Moreover, a novel approach to identify gestures within activities is here presented. The performance of the proposed methodology is tested on the two heterogeneous datasets related to cross-country skiing and daily activities. 

With the aim of monitoring the human activity, wearable devices provide an enhanced usability and a seamless human experience with respect to other portable devices (e.g. smartphones) in critical tasks as well as in leisure and sport activities. At the same time, though, wearable devices are more resource-constrained in terms of computational capability and memory, which calls for the design of algorithmic solutions that explicitly take into account these issues. In this paper, a parsimonious approach for activity recognition with wearable devices is presented. The methodology is based on Relevant Vector Machines (RVMs), a sparse machine learning framework for classification, and allows to tackle the activity recognition problem by identifying the two phases of Event Identification and Gesture Recognition. The performance of the presented methodology is tested on the interesting case study of cross- country skiing (classic style): such a dataset presents three different classes of gestures in addition to non-gesture activities and has been obtained by recording the training sessions of a heterogeneous set of executors in different environment conditions.

In the context of activity recognition, wearable devices are nowadays the preferable hardware thanks to their usability, user experience and performances; at the same time, these devices present limitations in terms of computational capability and memory, which force the algorithm design to be at the same time efficient and simple. In this work, we adopt Symbolic Aggregate Approximation (SAX), a symbolic approach for information retrieval in time series data that allows dimensionality and numerosity reduction; SAX is employed here, in combination with 1-Nearest Neighbor classifier, to identify activity phases in continuous repetitive activities from inertial time-series data. The proposed approach is validated on a cross-country skiing dataset and on a daily living activities dataset. 

In this paper, a Machine Learning (ML) approach is presented that exploits accelerometers data to deal with gesture recognition (GR) problems. The proposed methodology aims at providing high accuracy classi?cation for Home Automation systems, which are generally user independent, device independentand device orientation independent, an heterogeneous scenario that was not fully investigated in previous GR literature. The approach illustrated in this work is composed of three main steps: event identi?cation, feature extraction and ML-based classi?cation; elements of novelty of the proposed approach are:

1. a pre-processing phase based on Principal Component Analysis to increase the performance in real-world scenario conditions;

2.the development of parsimonious novel classi?cation techniques based on Sparse Bayesian Learning.

This methodology is tested on two datasets of 4 gesture classes (horizontal, vertical, circles and eight-shaped movements) and on a further dataset with 8 classes. In order to authentically describe a real-world Home Automation environment, the gesture movements are collected from more than 30 people who freely perform any gesture. It results a dictionary of 12 and 20 different movements respectivelyin the case of the 4-class and the 8-class databases. 

The interaction based on gestures has become a prominent approach to interact with electronic devices. In this paper a Machine Learning (ML) based approach to gesture recognition (GR) is illustrated; the proposed tool is freestanding from user, device and device orientation. The tool has been tested on a heterogeneous dataset representative of a typical application of gesture recognition. In the present work two novel ML algorithms based on Sparse Bayesian Learning are tested versus other classification approaches already employed in literature (Support Vector Machine, Relevance Vector Machine, k-Nearest Neighbor, Discriminant Analysis). A second element of novelty is represented by a Principal Component Analysis- based approach, called Pre-PCA, that is shown to enhance gesture recognition with heterogeneous working conditions. Feature extraction techniques are also investigated: a Principal Component Analysis based approach is compared to Frame- Based Description methods.