Lifelogging, as a growing interest, is a term referring to people digitally capturing all the information produced by them in daily life. Lifelog is a data collection of records of an individual's daily activities in one or more media forms. In this thesis, we collect lifelog data by using a mobile phone or a Microsoft Research SenseCam worn around subjects' necks during their daily life. We then propose a way to organize the lifelog data - a metric based model for event segmentation. Further more, we analyse the data properties through constructing the experience graphs from the recorded images. This thesis involves two parts, the details are as follows:
Firstly, we describe a metric-based model for event segmentation of sensor data recorded by a mobile phone worn around subjects' necks during their daily life. More specifically, we aim at detecting human daily event boundaries by analysing the recorded triaxial accelerometer signals and images sequence (lifelog data). In the experiments, different signal representations and three boundary detection models are evaluated on a corpus of 2 subjects over total 24 days. The contribution of this work is three-fold. First, we find that using accelerometer signals can provide much more reliable and significantly better performance than using image signals with MPEG-7 low level features. Second, the models using the accelerometer data based on the world's coordinates system can provide equally or even much better performance than using the accelerometer data based on the device's coordinates system. Finally, our proposed model has a better performance than the state of the art system.
Secondly, we investigate data obtained from subjects wearing a Microsoft Research SenseCam as they engaged in their every day activities. We construct experience graphs for each subject from their corresponding images by using two different image representation methods - color histogram and color correlogram. The statistical analyses of these graphs show that they have a small world structure which is characterized by high proximity ratios and sparse connectivity, independent of the representation used. However, the degree distribution analyses show that they are not scale-free, broad-scale or even single-scale. Furthermore, we also find that the graphs constructed based on the color correlogram representation, which is better than the color histogram in many content-based image retrieval systems, have shorter average path lengths and higher global clustering coefficients than the graphs constructed based on the color histogram representation.