Recognizing Human Activity in Free-Living Using Multiple Body-Worn Accelerometers

Abstract

Objectives: Recognising human activity is very useful for an investigator about a patient's behaviour and can aid in prescribing activity in future recommendations. The use of body worn accelerometers has been demonstrated to be an accurate measure of human activity, however research looking at the use of multiple body worn accelerometers in a free living environment to recognise a wide range of activities is not evident. This study aimed to successfully recognise activity and subcategory activity types through the use of multiple body worn accelerometers in a free living environment. Method: Ten participants (Age = 23.1 ± 1.7 years, height =171.0 ± 4.7 cm, mass = 78.2 ± 12.5 Kg) wore nine body-worn accelerometers for a day of free living. Activity type was identified through the use of a wearable camera, and sub category activities were quantified through a combination of freeliving and controlled testing. A variety of machine learning techniques consisting of pre-processing algorithms, feature and classifier selections were tested, accuracy and computing time were reported. Results: A fine k-nearest neighbour classifier with mean and standard deviation features of unfiltered data reported a recognition accuracy of 97.6%. Controlled and free-living testing provided highly accurate recognition for sub-category activities (>95.0%). Decision tree classifiers and maximum features demonstrated to have the lowest computing time. Conclusions: Results show recognition of activity and subcategory activity types is possible in a free living environment through the use of multiple body worn accelerometers. This method can aid in prescribing recommendations for activity and sedentary periods for healthy living.

Simultaneous Bayesian Recognition of Locomotion and Gait Phases With Wearable Sensors

“…These methods achieved accuracies between 77.3% and 99%, however, they need a large number of sensors that also limits these works to indoor applications [26], [27]. SVM and knearest neighbour algorithms, together with 9 accelerometers distributed from the torso to the ankle, achieved an accuracy of 97.6% for recognition of ADLs [28]. The combination of plantar pressure sensors with multi-class SVMs allowed the recognition of normal walking, stair ascent and stair descent activities with accuracies between 91.9% and 95.2% [29].…”

Towards Human Motion Tracking: Multi-Sensory IMU/TOA Fusion Method and Fundamental Limits

“…Currently, WS-based HMT systems [6] are mainly composed of inertial sensors, such as accelerometers and gyroscopes. However, inertial sensors may inevitably throw off errors that accumulate over time [8,15,25]. The accumulative and drift error is the biggest challenge faced by WS-based HMT systems.…”

“…As human motion tracking could be viewed as a multiple-target localization issue of human body joints, tracking accuracy is the most important consideration. However, sensor drift errors and distortion (especially in long time monitoring) are the main problems [8,15,25].…”

“…In essence, human motion tracking can be viewed as a local multi-target real-time high-precision three-dimensional positioning problem [23]. Ultra-Wideband (UWB)-based time-of-arrival (TOA) ranging is the most commonly used high-precision localization technology, its measurement accuracy can reach the centimeter level, and it does not have the drawback of accumulative errors, compared with inertial-sensor-based HMT systems [23,25]. The size of TOA chip is small enough to be integrated into wearable devices.…”

“…Some achievements on IMU/TOA fusion have been reported in many studies (e.g., [20,21,23,25]) However, state-of-the-art studies, (e.g., [23,27]), mainly face the following two drawbacks:…”