Multi-sensor activity recognition of an elderly person.

Abstract

The rapid increase in the number of ageing population brings major issues to health care including a rise in care cost, high demand in long- term care, burden to caregivers, and insufficient and ineffective care. Activity recognition can be used as the key part of the intelligent sys- tems to allow elderly people to live independently at homes, reduce care cost and burden to the caregivers, provide assurance for the fam- ilies, and promote better care. However, current activity recognition systems mainly focus on the technical aspect i.e. systems accuracy and neglects the practical aspects such as acceptance, usability, cost and privacy. The practicality of the system is the vital indication whether the system will be adopted. This research aims to develop the activity recognition system which considers both practical and technical aspects using multiple wrist-worn sensors. An extensive literature review in wearable sensor based activity recog- nition and its applications in healthcare have been carried out. Novel multi-sensor activity recognition utilising multiple low-cost, non-intrusive, non-visual wearable sensors is proposed. The sensor fusion is per- formed at feature and classi er levels using the proposed feature se- lection and classi er combination techniques. The multi-sensor ac- tivity recognition data sets have been collected. The rst data set contains data from accelerometer collected from seven young adults. The second data set contains data from accelerometer, altimeter, and temperature sensor collected from 12 elderly people in home environ- ment performing 10 activities. The third data set contains sensor data from accelerometer, gyroscope, temperature sensor, altimeter, barometer, and light sensor worn on the users wrist and a heart rate monitor worn over the users chest. The data set is collected from 12 elderly persons in a real home environment performing 13 activities. This research proposes two feature selection methods, Feature Com- bination (FC) and Maximal Relevancy and Maximal Complementary (MRMC), based on the relationship between feature and classes as well as the relationship between a group of features and classes. The experimental studies show that the proposed techniques can select an optimum set of features from irrelevant, overlapped, and partly over- lapped features. The studies also show that FC and MRMC obtain higher classi cation performances than popular techniques including MRMR, NMIFS, and Clamping. Two classi er combination tech- niques based on Genetic Algorithm (GA) are proposed. The rst technique called GA based Fusion Weight (GAFW), uses GA nd the optimum fusion weights. The results indicate that 99% of classi er fusion using GAFW achieves equal or higher accuracy than using only the best classi er. While other fusion weight techniques cannot guar- antee accuracy improvement, GAFW is a more suitable method for determining fusion weight regardless which fusion techniques are used. Another algorithm called GA based Combination Model (GACM) is proposed to nd the optimal combination between classi er, weight function, and classi er combiners. The algorithm does not only nd the model which has the minimum classi cation error but also select the one that is simpler. Other criteria e.g. select the classi er with low computation can also be easily added to the algorithm. The re- sults show that in general GACM can nd the optimum combinations automatically. The comparison against manually selection revealed that there is no statistical signi cant in the performances. Applications of the proposed work in home care and decision support system are discussed The results of this research will have a signi cant impact on the future health care where people can be health monitored from their homes to promote healthy living, detect any changes in behaviour, and improve quality of care.