Designing Frameworks for Recognizing HUMAN ACTIVITIES Using Machine Learning Techniques

Daily Activity Recognition has profound importance in our lives for applications

such as, behavior analysis at indoor, smart healthcare, entertainment and

surveillance applications. Smartphone based Human Activity Recognition (HAR)

frameworks provide a convenient and cost eective solution to the problems. Seamless

monitoring of elderly people living alone mostly at indoor has become feasible

applying smartphone based HAR with assured privacy compared to camera based

monitoring. Machine learning techniques are adopted as the problem needs to scale

well with varying datasets and conditions. The main challenges behind such a successful

HAR framework for smartphones include (i) stable recognition performance

irrespective of dierent hardware conguration and usage behavior of smartphones

and (ii) annotating data in real-life scenario, especially for activity transitions involving

composite activities. These are the main challenges addressed in this thesis.

The rst challenge is addressed in mainly two ways- (i) through designing ensemble

of classiers and (ii) fusion of sensors. Fusion of dierent smartphone sensors

and fusion of wearable sensing with smartphone sensors are explored for detailed

activity recognition. In order to address the second challenge of grossly labeled

datasets, Multiple Instance Multiple Label (MIML) learning methods are explored.

An ensemble of MIML-KNN classiers is designed that is found to predict activity

sequence as well as set of activity transitions performed for a given time period

with considerable accuracy. The framework is also found to detect unknown activity

combinations either partially or totally.