Evaluation of accelerometer based multi-sensor versus single-sensor activity recognition systems.

摘要: Abstract Physical activity has a positive impact on people's well-being and it had been shown to decrease the occurrence of chronic diseases in older adult population. To date, substantial amount research studies exist, which focus recognition using inertial sensors. Many these adopt single sensor approach proposing novel features combined with complex classifiers improve overall accuracy. In addition, implementation advanced feature extraction algorithms exceed computing ability most current wearable platforms. This paper proposes method multiple sensors distributed body locations overcome this problem. The objective proposed system is achieve higher accuracy “light-weight” signal processing algorithms, run based comprised computationally efficient nodes. For analysing evaluating multi-sensor system, eight subjects were recruited perform normal scripted activities different life scenarios, each repeated three times. Thus total 192 recorded resulting 864 separate annotated states. methods for designing such required consideration following: pre-processing sampling rate, selection classifier selection. Each investigated appropriate selected trade-off between execution time. A comparison six systems, employ or sensors, presented. experimental results illustrate that can an 96.4% by adopting mean variance features, Decision Tree classifier. demonstrate elaborate sets are not high accuracies system.