Xavier Parra

Human activity recognition on smartphones using a multiclass hardware-friendly support vector machine

Activity-Based Computing [1] aims to capture the state of the user and its environment by exploiting heterogeneous sensors in order to provide adaptation to exogenous computing resources. When these sensors are attached to the subjects body, they permit continuous monitoring of numerous physiological signals. This has appealing use in healthcare applications, e.g. the exploitation of Ambient Intelligence (AmI) in daily activity monitoring for elderly people. In this paper, we present a system for human physical Activity Recognition (AR) using smartphone inertial sensors. As these mobile phones are limited in terms of energy and computing power, we propose a novel hardware-friendly approach for multiclass classification. This method adapts the standard Support Vector Machine (SVM) and exploits fixed-point arithmetic for computational cost reduction. A comparison with the traditional SVM shows a significant improvement in terms of computational costs while maintaining similar accuracy, which can contribute to develop more sustainable systems for AmI.

Transition-Aware Human Activity Recognition Using Smartphones

This work presents the Transition-Aware Human Activity Recognition (TAHAR) system architecture for the recognition of physical activities using smartphones. It targets real-time classification with a collection of inertial sensors while addressing issues regarding the occurrence of transitions between activities and unknown activities to the learning algorithm. We propose two implementations of the architecture which differ in their prediction technique as they deal with transitions either by directly learning them or by considering them as unknown activities. This is accomplished by combining the probabilistic output of consecutive activity predictions of a Support Vector Machine (SVM) with a heuristic filtering approach. The architecture is validated over three case studies that involve data from people performing a broad spectrum of activities (up to 33), while carrying smartphones or wearable sensors. Results show that TAHAR outperforms state-of-the-art baseline works and reveal the main advantages of the architecture.

K-SVCR. A support vector machine for multi-class classification

Abstract The problem of multi-class classification is usually solved by a decomposing and reconstruction procedure when two-class decision machines are implied. During the decomposing phase, training data are partitioned into two classes in several manners and two-class learning machines are trained. To assign the class for a new entry, machines’ outputs are evaluated in a specific pulling scheme. This article introduces the “Support Vector Classification-Regression” machine for K -class classification purposes ( K -SVCR), a new training algorithm with ternary outputs {−1,0,+1} based on Vapniks Support Vector theory. This new machine evaluates all the training data into a 1-versus-1-versus-rest structure during the decomposing phase by using a mixed classification and regression SV Machine (SVM) formulation. For the reconstruction, a specific pulling scheme considering positive and negative votes has been designed, making the overall learning architecture more fault-tolerant as it will be demonstrated.

Human Activity Recognition on Smartphones with Awareness of Basic Activities and Postural Transitions

Postural Transitions (PTs) are transitory movements that describe the change of state from one static posture to another. In several Human Activity Recognition (HAR) systems, these transitions cannot be disregarded due to their noticeable incidence with respect to the duration of other Basic Activities (BAs). In this work, we propose an online smartphone-based HAR system which deals with the occurrence of postural transitions. If treated properly, the system accuracy improves by avoiding fluctuations in the classifier. The method consists of concurrently exploiting Support Vector Machines (SVMs) and temporal filters of activity probability estimations within a limited time window. We present the benefits of this approach through experiments over a HAR dataset which has been updated with PTs and made publicly available. We also show the new approach performs better than a previous baseline system, where PTs were not taken into account.

Training Computationally Efficient SmartphoneBased Human Activity Recognition Models

The exploitation of smartphones for Human Activity Recognition (HAR) has been an active research area in which the development of fast and efficient Machine Learning approaches is crucial for preserving battery life and reducing computational requirements. In this work, we present a HAR system which incorporates smartphone-embedded inertial sensors and uses Support Vector Machines (SVM) for the classification of Activities of Daily Living (ADL). By exploiting a publicly available benchmark HAR dataset, we show the benefits of adding smartphones gyroscope signals into the recognition system against the traditional accelerometer-based approach, and explore two feature selection mechanisms for allowing a radically faster recognition: the utilization of exclusively time domain features and the adaptation of the L1 SVM model which performs comparably to non-linear approaches while neglecting a large number of non-informative features.

Fault tolerance in the learning algorithm of radial basis function networks

A method of supervised learning is described which improves fault tolerance by means of modifying the learning algorithm in order to introduce significant information related to fault tolerance during training. The method exploits the evolutive nature of the learning algorithm of radial basis function networks and employs optimisation techniques to control the balance between generalisation performance and fault tolerance. The technique developed is specific to the neural architecture employed, though it can be used concurrently with other more traditional approaches like training with faults or retraining. The fault-tolerant algorithm presented provides a simple and efficient means of improving fault tolerance, and this is illustrated using examples taken from two different classification problems.

Robust multi-dimensional motion features for first-person vision activity recognition

We design a set of multi-dimensional motion features from first-person video.We extract virtual inertial data from video only.We combine motion magnitude, direction and dynamics with virtual inertial data.The features are independent of the classifier and validated on multiple datasets.Two new datasets are made available to the research community. We propose robust multi-dimensional motion features for human activity recognition from first-person videos. The proposed features encode information about motion magnitude, direction and variation, and combine them with virtual inertial data generated from the video itself. The use of grid flow representation, per-frame normalization and temporal feature accumulation enhances the robustness of our new representation. Results on multiple datasets demonstrate that the proposed feature representation outperforms existing motion features, and importantly it does so independently of the classifier. Moreover, the proposed multi-dimensional motion features are general enough to make them suitable for vision tasks beyond those related to wearable cameras.

Ambulatory Mobility Characterization Using Body Inertial Systems: An Application to Fall Detection

The aim of this paper is to study the use of a prototype of wearable device for long term monitoring of gait and balance using inertial sensors. First, it is focused on the design of the device that can be used all day during the patient daily life activities, because it is small, usable and non invasive. Secondly, we present the system calibration to ensure the quality of the sensors data. Afterwodrs, we focus in the experimental methodology for data harvest from extensive types of falls. Finally a statistical analysis allows us to determine the discriminant information to detect falls.

Gyro-based Camera-motion Detection in User-generated Videos

We propose a gyro-based camera-motion detection method for videos captured with smartphones. First, the delay between the acquisition of video and gyroscope data is estimated using similarities induced by camera motion in the two sensor modalities. Pan, tilt and shake are then detected using the dominant motions and high frequencies in the gyroscope data. Morphological operations are applied to remove outliers and to identify segments with continuous camera-motion. We compare the proposed method with existing methods that use visual or inertial sensor data.

Generalisation Improvement of Radial Basis Function Networks Based on Qualitative Input Conditioning for Financial Credit Risk Prediction

The rating is a qualified assessment about the credit risk of bonds issued by a government or a company. There are specialised rating agencies, which classify firms according to their level of risk. These agencies use both quantitative and qualitative information to assign ratings to issues. The final rating is the judgement of the agencys analysts and reflects the probability of issuer default. Since the final rating has a strong dependency on the experts knowledge, it seems reasonable the application of learning based techniques to acquire that knowledge. The learning techniques applied are neural networks and the architecture used corresponds to radial basis function neural networks. A convenient adaptation of the variables involved in the problem is strongly recommended when using learning techniques. The paper aims at conditioning the input information in order to enhance the neural network generalisation by adding qualitative expert information on orders of magnitude. An example of this method applied to some industrial firms is given.