A data fusion-based hybrid sensory system for older people’s daily activity recognition.

Abstract

Population aged 60 and over is growing faster. Ageing-caused changes, such as physical or cognitive decline, could affect people’s quality of life, resulting in injuries, mental health or the lack of physical activity. Sensor-based human activity recognition (HAR) has become one of the most promising assistive technologies for older people’s daily life. Literature in HAR suggests that each sensor modality has its strengths and limitations and single sensor modalities may not cope with complex situations in practice. This research aims to design and implement a hybrid sensory HAR system to provide more comprehensive, practical and accurate surveillance for older people to assist them living independently. This reseach: 1) designs and develops a hybrid HAR system which provides a spatio- temporal surveillance system for older people by combining the wrist-worn sensors and the room-mounted ambient sensors (passive infrared); the wearable data are used to recognize the defined specific daily activities, and the ambient information is used to infer the occupant’s room-level daily routine; 2): proposes a unique and effective data fusion method to hybridize the two-source sensory data, in which the captured room-level location information from the ambient sensors is also utilized to trigger the sub classification models pretrained by room-assigned wearable data; 3): implements augmented features which are extracted from the attitude angles of the wearable device and explores the contribution of the new features to HAR; 4:) proposes a feature selection (FS) method in the view of kernel canonical correlation analysis (KCCA) to maximize the relevance between the feature candidate and the target class labels and simultaneously minimizes the joint redundancy between the already selected features and the feature candidate, named mRMJR-KCCA; 5:) demonstrates all the proposed methods above with the ground-truth data collected from recruited participants in home settings. The proposed system has three function modes: 1) the pure wearable sensing mode (the whole classification model) which can identify all the defined specific daily activities together and function alone when the ambient sensing fails; 2) the pure ambient sensing mode which can deliver the occupant’s room-level daily routine without wearable sensing; and 3) the data fusion mode (room-based sub classification mode) which provides a more comprehensive and accurate surveillance HAR when both the wearable sensing and ambient sensing function properly. The research also applies the mutual information (MI)-based FS methods for feature selection, Support Vector Machine (SVM) and Random Forest (RF) for classification. The experimental results demonstrate that the proposed hybrid sensory system improves the recognition accuracy to 98.96% after applying data fusion using Random Forest (RF) classification and mRMJR-KCCA feature selection. Furthermore, the improved results are achieved with a much smaller number of features compared with the scenario of recognizing all the defined activities using wearable data alone. The research work conducted in the thesis is unique, which is not directly compared with others since there are few other similar existing works in terms of the proposed data fusion method and the introduced new feature set.