Improvement of Emotional Video Scene Retrieval System for Lifelog Videos Based on Facial Expression Intensity

Lifelog has been proposed, in which various data of daily life are acquired and accumulated, and utilized later. However, it is a problem that we can not immediately retrieve the necessary data from a large amount of accumulated data, so the lifelog data are not effectively used. This paper deals with lifelog videos. In order to make it easy to search the scene that the user wants to watch from the lifelog videos, Morikuni tried to construct a system that could search the scene considered to be important with a change in facial expression of the person and to present it in an easy-to-understand manner. After that, "facial expression intensity" which is a numerical representation of facial expressions was devised, and Maeda designed and constructed a video scene retrieval system for lifelog videos based on the facial expression intensity. In this paper, we aim to improve the user interface of this retrieval system and establish a method to estimate the threshold values of the facial expression intensity level. We propose and implement a method to calculate the threshold values using the k-means clustering. We compare the performance of the threshold values with the threshold values of the previous method, and show that the performance was improved.