Ortam destekli yaşam için kesintisiz atalet verisi kullanarak insan aktivitelerinin tanınması

Ambient intelligence systems open up numerous prospects for providing assistance to elderly and disabled people as well as healthy individuals. Wearable computing devices embroidered with data communication capabilities participating in IoT applications boost coverage of those systems, integrating intelligent modules to daily lives of people. Assistance offered by the intelligent systems can emerge as various use cases including smart healthcare embedded in daily living environments. Such systems incorporate activity recognition units as an integral part to determine the corresponding actuation. For activity detection units to get blended in the course of daily routines of people, detection of individual activities is not sufficient where additional constraints should also be considered alongside activity detection problem. These constraints necessitate dynamic segmentation of the streaming activity data during prediction stage in addition to activity detection. This thesis presents a spotting method considering a group of constraints within the scope of intelligent healthcare domain in daily living settings utilizing wearable sensor technology. This thesis contributes to improving the commonly used segmentation approaches in the sense that the spotting technique introduced in this thesis addresses several constraints which have not been tackled in previous work. Spotting the target activity in a sequence of activities and transitions without applying a pre-determined window size is a challenging task. The sliding window method, which is the typical approach in segmenting the continuous data stream, operates with optimal segment sizes chosen considering the type and duration of the activities. Nevertheless, activity type and duration are usually unknown to the detection unit in practice. In this thesis, we proposed Non-predetermined Size Windowing (NSW) scheme to spot the target activity performed in a sequence of unseen activities. NSW is built on classifying progress based features in multi-layer training and prediction stages where time domain progress is xpressed in terms of polynomials. Thus, it operates without incorporating the information regarding duration and type of the activities. We verified our method with a use case where data are acquired by a single wrist-worn 3D accelerometer. We compared our method against fixed size windowing performed with varying window sizes and feature extraction schemes; windowed energy, peak frequency, Shannon entropy and wavelet entropy. Additionally, we propose an Artificial Data Generation (ADG) scheme and further investigate its contribution in monitoring the adherence of the patients to the elbow flexion and extension physiotherapy exercise routines in their daily environments. Performing detection of an activity in the existence of unseen classes is a case of classification where multiclass classification schemes are inadequate since in some use cases acquisition of only target activity data is possible before prediction stage. One-class classification (OCC) is a strategy to deal with such use cases. As an alternative to adapting the architecture of an existing classifier to design an OCC scheme, ADG approach emerges as a practical way of performing OCC. The ADG solution utilized in this thesis is evaluated in a setting with a single wrist-worn accelerometer where training data are collected for the target class only and data that represent all other classes are produced artificially from the target activity data. We analyzed the ADG method coupling with four different classifiers; KNN, SVM, logistic regression and Naive Bayes classifiers. Finally, we quantified the relation between activity duration, size of activity set and success ratio to provide an insight for selecting the methods which segmentation schemes are built on top of. In a real time execution environment, utilization of a segmentation algorithm leads to partial availability of activity data until data corresponding to a particular activity or segment are completely acquired, which brings about the problem of incompleteness of data. Hence, a segmentation scheme should be resilient to such incompleteness, which is incompleteness mitigation (IM) property in a knowledge discovery scheme. Quantification of relation between activity duration, size of activity set and success ratio is an indication of degree of IM. We developed an Intraclass Correlation Coefficient (ICC) based metric to measure degree of IM. We analyzed peak frequency and wavelet entropy feature extraction schemes with KNN in an iterative setting to observe their capability in mitigating incompleteness of the acceleration data in several hand oriented activities.