Real time activity monitoring

Activity monitoring systems (AMS) are responsible for detecting actions performed by humans. For AMS to be effectively deployed in daily life, they should be operating in real-time and partition the continuously streaming activity data to determine what activity corresponds to each partition. In this thesis, a Support Vector Machine based real time continuous activity monitoring system, named RT-CAM, is proposed. We approach continuous activity detection problem by modelling the activities as simple and composite actions. Simple actions being the smallest meaningful actions which can not be further divided into smaller logical actions whereas composite actions are combinations of simple actions. The proposed model detects simple and composite activities in real time, collecting the data with a single 3D accelerometer to produce a non-invasive solution. We verified our model on hand oriented set of simple actions eat, pour, drink, toothBrush and turnKey, with real data acquired from human subjects instead of computer generated synthetic data. We showed that the selected activities can be distinguished in real time though they generate quite similar patterns to each other. The strength and novelty of the proposed model lies in the fact that the system does not necessitate being trained with patterns of transitions and does not run a dedicated algorithm for transition detection. We carried out experiments on 4 different subjects and present our best achieved results. Intra-person test results are the following: ToothBrush, drink, drink_toothBrush, toothBrush_drink and toothBrush_pour are recognized with 100% accuracy. Drink_toothBrush_pour and toothBrush_drink_pour_turnKey are detected with 80% and 70% accuracy respectively. Inter-person test results are the following: ToothBrush, drink and pour are detected with 100% accuracy. Drink_toothBrush, drink_toothBrush_pour and drink_toothBrush_turnKey are recognized with 80% accuracy. Real time overhead introduced by RT-CAM is 0.055 seconds, which is better than best achieved result in the literature. Considering all these features, RT-CAM is an applicable solution in real time continuous activity monitoring.