Introduction: Of particular interest in healthcare monitoring, automatic recognition of human activity has become an extensively researched topic. Unlike the majority of recognition systems, our study focuses on the classification of dynamic and stationary activities extended to natural out-of-lab conditions, with a two-fold purpose: (i) find the optimal sensors placements on the body and (ii) assess the effect of combining the heart rate (HR) information to the acceleration. Methods: The dataset was established on 8 volunteers freely performing 5 activities: run, walk, cycle, car ride, and rest (stand or sit). 4 nodes manufactured by the SHIMMER Company (Dublin, Ireland) recording 3D raw acceleration signals (in gravity unit: g) were attached to the ankle, chest, wrist and hip of the subjects and 6 electrodes were placed on the chest to extract the HR measures from the RR waves of the recorded Electrocardiogram (ECG) signals. The recognition model is based on a previously developed method (AbdulRahman et al., 2015) that extracts the spectral distances features from acceleration data. These features are then used as input for 3 commonly used classifiers: the K-Nearest neighbors (KNN), the Naive Bayes (NB) and the Decision Tree (DT) for activity prediction (Lara et al. 2013). Results: With the 4-position accelerometers, an overall classification rate of 99.0% ± 0.4 was achieved by the KNN classifier, while the NB and the DT classifiers marked 87.8% ± 1.1 and 95.9% ± 1.1 respectively. 89.8% ± 1.0, 94.5% ± 0.8, 95.4% ± 0.7 and 95.7% ± 0.7 are the system responses when considering the single node on the chest, the wrist, the hip and the ankle respectively. These accuracies raise to 94.1% ± 1.1, 97.9% ± 0.6, 98.7% ± 0.5, and 98.7% ± 0.3 when coupling the HR information to each corresponding position. Discussion: The lower limbs sensors hold the most valuable information for the prediction of the dynamic and stationary activity types. Unlike the `rest` and `car ride` activities, `run` and `walk` are correctly detected, which support the idea that ambulation behaviors involve quasi-periodic movement of the body (Altun et al. 2010). Combining acceleration and HR data enhance the accuracy of the identification. Thus, with an HR monitor and only a single unit accelerometer, it is accurately sufficient to classify dynamic and stationary activities (mean accuracy higher than 97%).
© Copyright 2016 21st Annual Congress of the European College of Sport Science (ECSS), Vienna, 6. -9. July 2016. Veröffentlicht von University of Vienna. Alle Rechte vorbehalten.