Characterizing everyday activities from visual lifelogs based on enhancing concept representation

摘要

The proliferation of wearable visual recording devices such as SenseCam, Google Glass, etc. is creating opportunities for automatic analysis and usage of digitally-recorded everyday behavior, known as visual lifelogs. Such information can be recorded in order to identify human activities and build applications that support assistive living and enhance the human experience. Although the automatic detection of semantic concepts from images within a single, narrow, domain has now reached a usable performance level, in visual lifelogging a wide range of everyday concepts are captured by the imagery which vary enormously from one subject to another. This challenges the performance of automatic concept detection and the identification of human activities because visual lifelogs will have such variety of semantic concepts across individual subjects. In this paper, we characterize the everyday activities and behavior of subjects by applying a hidden conditional random field (HCRF) algorithm on an enhanced representation of semantic concepts appearing in visual lifelogs. This is carried out by first extracting latent features of concept occurrences based on weighted non-negative tensor factorization (WNTF) to exploit temporal patterns of concept occurrence. These results are then input to an HCRF-based model to provide an automatic annotation of activity sequences from a visual lifelog. Results for this are demonstrated in experiments to show the efficacy of our algorithm in improving the accuracy of characterizing everyday activities from individual lifelogs. The overall contribution is a demonstration that using images taken by wearable cameras we can capture and characterize everyday behavior with a level of accuracy that allows useful applications which measure, or change that behavior, to be developed.