SVM-Based Multimodal Classification of Activities of Daily Living in Health Smart Homes: Sensors, Algorithms, and First Experimental Results

Abstract

Abstract-By 2050, about a third of the French population will be over 65. Our laboratory's current research focuses on the monitoring of elderly people at home, to detect a loss of autonomy as early as possible. Our aim is to quantify criteria such as the international ADL or the French AGGIR scales, by automatically classifying the different Activities of Daily Living performed by the subject during the day. A Health Smart Home is used for this. Our Health Smart Home includes, in a real flat, Infra-Red Presence Sensors (location), door contacts (to control the use of some facilities), temperature and hygrometry sensor in the bathroom, and microphones (sound classification and speech recognition). A wearable kinematic sensor also informs on postural transitions (using pattern recognition) and walk periods (frequency analysis). This data collected from the various sensors, is then used to classify each temporal frame into one of the activities of daily living that was previously acquired (seven activities: hygiene, toilet use, eating, resting, sleeping, communication, and dressing/undressing). This is done using Support Vector Machines. We performed a one-hour experimentation with 13 young and healthy subjects to determine the models of the different activities and then we tested the classification algorithm (cross-validation) with real data.

Activity recognition in smart homes with self verification of assignments

“…ARSH-SV is evaluated through a comprehensive performance evaluation metrics containing seven performance measures using nine publicly available smart home datasets. The results are compared with the state-of-the-art activity recognition approaches [13,15,30,35] using three-fold cross validation. The recognition rate (number of correctly recognized instances out of all instances in the test set) and F1-score (Eq.…”

“…Feature selection techniques have been applied in the activity recognition problem to select significant and discriminant subset of features [13,27,31,44]. Minimum Redundancy Maximum Relevance (mRMR) can be used to select the best feature subset for target classes, then SVM is used for the classification of activities [27].…”

“…The mRMR and Information Gain (IG) methods are applied for feature selection and then four classifiers SVM, LogitBoost, Bayes Belief Networks and ANN are compared for recognition of activities [31], where LogitBoost proves to be better among the four due to its ability of generating a stronger classifier by combining the several weak classifiers. In a PCA-SVM approach [13], PCA is used to extract the significant features and the multi-class SVM (one-versus-one) is applied for classification of activities. Using SVM for classification, the comparison of the feature selection methods, filter and wrapper based on single and sequential feature selection shows that wrapper method based on sequential feature selection performs better [44], since it also considers the redundancy of features during the selection process, unlike the other two.…”

“…Activity segmentation is performed using different techniques, sliding windows [7], relative weighting of objects in adjacent activities [8] or pattern mining [9], just to name a few. Segmented activity instances are classified in activity classes using different learning models such as Hidden Markov Model (HMM) [10], Conditional Random Fields (CRF) [11], Naive Bayes (NB) [12], Support Vector Machine (SVM) [13], Artificial Neural Network (ANN) [14,15], and Decision Tree (DT) [16]. In activity classification, a false assignment could occur due to the unreliable nature of sensor data [17], incorrect execution of an activity [18], similar activities due to overlapping in features [19] or inability of a learning algorithm to assign the correct label [20].…”

“…In health care systems, the reliability of activity recognition models is extremely important, therefore along with the correct recognition of activities, a model should also be capable of detecting and avoiding false assignments [20]. Most of the activity recognition approaches while focusing on the segmentation and recognition may ignore false assignments [10,13,21]. Additionally the existing approaches that exploit the temporal pattern for recognition assume that the activities follow a certain predefined sequence of events, however a fixed sequence may not always be the case, since a single activity can be performed in different ways by different users [22].…”