Article Preview
TopIntroduction
According to the 2015 Telecommunications White Paper from the Japanese Ministry of Internal Affairs and Communications, the number of “things” related to the Internet of Things (IoT) is expected to increase tremendously; a wide variety of data will be collected and analyzed via IoT and movements to link these are expected to enhance work efficiency, among other things. Naturally, this includes the treatment of big data, and when analyzing data collected in such large quantities, their analysis could be performed interactively, allowing interpretations from many different angles. In particular, security camera images, while considering privacy, are linked to a large volume of data collected using IoT. The reason for this is exceptionally simple. For example, to monitor the daily behavioral patterns of elderly people and children, it is convenient to construct surveillance systems using data collected from IoT because it is portable and allows us to analyze and extract data in order to evaluate whether their behaviors have changed or are in danger of doing so. There are the researches in the computer network for the realization of such a system (Shintaro, Mariko, Mingrui, Yoshikazu, & Toshimitsu, 2013). However, with this simple idea, the main point is how to construct a system in which the behavior pattern data collected from IoT can be linked to actual behavior. Although the search engine to connect sensor data with an event has been already proposed (Koji, Yutaka, Takuya, Yasue, Yasushi, & Takeshi, 2010; Takeshi, Yasue, Takuya, Koji, Yutaka, & Yasushi, 2010), we believe that if a video is captured at the same time as the IoT data are collected, specification of these data is a simple matter.
In relation to the treatment of images, deep learning using deep neural networks (DNNs) has attracted a lot of attention in recent years because it successfully demonstrated automatic organization of neurons selectively reacting to images of a cat (Quoc,2013) and high recognition performance in relation to difficult tasks, such as general object recognition. In other words, this success engendered several methods to mitigate model complexity including weight decay (Stephen, & Lorien, 1989) and weight sharing (Le, Bernhard, John, Donnie, Richard, Wayne, & Lawrence, 1990; Steven, & Geoffrey, 1992) and to roughly form DNN at an early stage through prior learning (Geoffrey, Simon, & Yee-Whye, 2006). Furthermore, by creating an ensemble of multiple neural networks, such as DropOut (Geoffrey, Nitish, Alex, Ilya, & Ruslan, 2012) or DropConnect (Li, Matthew, Sixin, Yann, & Rob, 2013), which is a model assumed to have information deficiency in it, it has greatly contributed to the development of various parameter tuning for keeping generality.
Analyzing the action parameters from videos captured at the same time as data is collected from IoT, it is clear that although human eyes can specify actions easily by directly viewing them, due to issues such as their processing ability and fatigue, they soon encounter limitations when processing large quantities for long periods. Therefore, there is a demand for the automation of action pattern analysis using such images, and DNN can be used to achieve this.
In DNN image processing, if a convolutional neural network (CNN) (Alex, Ilya, & Geoffrey, 2012) or sparse coding is used, the response neurons corresponding to image input in the feature extraction layer learn to respond and become able to recognize objects. By utilizing this process on a video frame by frame, it is possible to consider a discrete feature vector corresponding to this video (Yasuyuki, Takashi, & Kuniaki, 2015). However, we have considered utilizing a continuous trajectory of an object or person moving consecutively. These trajectories are formed from a nonlinear mapping into a low-dimensional feature space by handling feed-forward neural networks without CNN or sparse coding.