Evaluation of a smartphone human activity recognition application with able-bodied and stroke participants

Abstract

BackgroundMobile health monitoring using wearable sensors is a growing area of interest. As the world’s population ages and locomotor capabilities decrease, the ability to report on a person’s mobility activities outside a hospital setting becomes a valuable tool for clinical decision-making and evaluating healthcare interventions. Smartphones are omnipresent in society and offer convenient and suitable sensors for mobility monitoring applications. To enhance our understanding of human activity recognition (HAR) system performance for able-bodied and populations with gait deviations, this research evaluated a custom smartphone-based HAR classifier on fifteen able-bodied participants and fifteen participants who suffered a stroke.MethodsParticipants performed a consecutive series of mobility tasks and daily living activities while wearing a BlackBerry Z10 smartphone on their waist to collect accelerometer and gyroscope data. Five features were derived from the sensor data and used to classify participant activities (decision tree). Sensitivity, specificity and F-scores were calculated to evaluate HAR classifier performance.ResultsThe classifier performed well for both populations when differentiating mobile from immobile states (F-score > 94 %). As activity recognition complexity increased, HAR system sensitivity and specificity decreased for the stroke population, particularly when using information derived from participant posture to make classification decisions.ConclusionsHuman activity recognition using a smartphone based system can be accomplished for both able-bodied and stroke populations; however, an increase in activity classification complexity leads to a decrease in HAR performance with a stroke population. The study results can be used to guide smartphone HAR system development for populations with differing movement characteristics.

Smart Lifelogging: Recognizing Human Activities using PHASOR

“…Data mining approaches were utilized to build classifiers with an information theory based ranking of features as the pre-processing step. Recently, Capela et al in (Capela et al, 2016) proposed a new method that can take into account different types of users who have differences in walking biomechanics. This system is considered as more affordable-price and convenient solution than using wearable sensors.…”

“…The first component utilizes accelerometer, gyroscope, and barometer sensors to gather data from human activities. These sensors can be used alone (Siirtola and Roning, 2012) (Bayat et al, 2014), or combined together (Shoaib, 2013) (Chetty et al, 2015) (Capela et al, 2016). The second component is built by using different classification methods such as Support Vector Machine (SVM), k-Nearest Neighbour (k-NN/IBk), or others (Lara and Labrador, 2013) (Shoaib et al, 2015).…”

“…In general, most smart-phone based HAR systems are built with three major components: sensory data acquisition, model training, and activity recognition (Capela et al, 2016). The first component utilizes accelerometer, gyroscope, and barometer sensors to gather data from human activities.…”

“…Among types of wearable sensors, smartphones are preferred as the most convenient equipment that can monitor human activities because of its mobility, user-friendly interface, longtime attachment, and available resources such as various embedded sensors, strong CPU, memory, and battery (Shoaib et al, 2015). According to (Lara and Labrador, 2013) (Shoaib et al, 2015), existing challenges include in-sufficient (standard) training data (Vavoulas et al, 2016) (Ojetola et al, 2015), varying positions and orientations of smartphones on the human body (Miao et al, 2015), resource consumption and privacy (Siirtola and Roning, 2012), dynamic and adaptive sensor selection (Capela et al, 2016) and online versus offline training for classification methods (Shoaib et al, 2015)(Google Activity Recognition API, 2016), etc. Nevertheless, none of the related work discusses the human factor in Internet of Everything (IoE) systems.…”

Recognizing Complex Human Activities using Hybrid Feature Selections based on an Accelerometer Sensor

“…In each instance, this method finds the nearest hit (data point from the same class) and nearest misses (data point from a different class). These data are calculated by weighting them based on their relevance (Akhavian & Behzadan, 2015;Capela et al, 2016). Equation 1 below shows how to calculate the feature weight based on its relevance.…”