Toru Nakata

Recognizing Human Activities in Video by Multi-resolutional Optical Flows

A method to recognize human activities captured in video is proposed. The method classifies basic human body activities, such as walking, running, gymnastic exercises and others. Applying Burt-Adelson Pyramid approach, the system extracts useful features consisting of multi-resolutional optical flows. This paper also reports coarseness limit of spatial resolution of optical flow for activity recognition; optical flows of 8 sub-areas covering the human body area are minimum requirement for the recognition. Also, the experiment examines effective weighting of multi-resolutional feature components. These results on recognition of coarse video will be useful for designing surveillance camera system

Temporal segmentation and recognition of body motion data based on inter-limb correlation analysis

A method for segmentation and recognition of human body behavior data is proposed. Recognition of human body movements is getting larger interests in robotic research field, since robots must recognize human behavior in order to interact with human in the real world. In addition, there is demand for quantitative methods to analyze human body movements, since human body movements can be used as models of robot behaviors. The author proposes a scheme for human behavior recognition based on two process steps: analysis of movement correlations among limbs and temporal segmentation of motion data. Inter-limb movement correlations are widely observed in various behaviors and well represent contents of behavior, so it will be a universal feature value for general behavior. Observing changes of inter-limb correlations, we can segment motion capture data into temporal fragment of action units. Using this segmentation technique in an experiment, the system succeeded recognizing various types of human behavior efficiently.

Human pose recognition by memory-based hierarchical feature matching

A human posture recognition system based on a memory-based approach is studied. Human body images extracted by depth data are labeled and stored in a database together with compressed feature values consist of higher-order local correlation values and outline diameters. The compressed features are used to speed up the database search in a hierarchical manner. The system can classify human body postures into 6 categories in the experiment. The system was robust against change of humans and light condition.

Human finding and body property estimation by using floor segmentation and 3D labelling

This paper describes a human finding and body property estimation system, which is able to recognize the floor surface and humans from distance information obtained from a stereo camera. The system is comprised of two algorithms; random two point sampling method and, 3D labeling method. The random two point sampling method can detect three-dimensional planer surfaces so as to recognize floor surface from input depth map. The 3D labeling method can recognize humans from dense disparity map by omitting the floor detected in the previous procedure. Finally the system estimate body height from the depth map and uses this to estimate head and leg length from regression formula in an anthology of human body measurement.

Vulnerability of the conventional accessible CAPTCHA used by the White House and an alternative approach for visually impaired people

Many people with visual impairments complain about the poor accessibility of conventional CAPTCHA systems because the audio-style test is too difficult for humans. Even a U.S. governmental site, the “We the People” public website, was criticized for the same reason, and thus it implemented a more accessible quiz-based CAPTCHA system. However, this system is vulnerable to simple heuristics. In this study, we demonstrate the insecurity of this type of CAPTCHA system. We demonstrate that our solver program can beat the CAPTCHA with a success rate of over 99%. In addition, we propose a new verbal-style system to replace the quiz-based CAPTCHA. Our system synthesizes several sentences, which have different degrees of naturalness in terms of their contextual meaning, from a set of source documents using a flexible-order Markov chain. Only human users can perceive the difference in the semantics and select the most (or the least) meaningful option correctly. This test is implemented in a verbal style, which means that it is universally suitable for any type of perceptual channel. We implemented our proposed scheme and analyzed its security based on experiments.

An Accessible CAPTCHA System for People with Visual Disability – Generation of Human/Computer Distinguish Test with Documents on the Net

We propose a new scheme of CAPTCHA that does not become a perceptual barrier for disable people. Our CAPTCHA system generates the tests in verbal style, so its use is not limited in specific perceptual channels. The tests are composed of several phrases and there are two kinds of tests: Human users try to (1) distinguish a phrase of strange meaning from others, and (2) identify the common topic among them.In our test we utilize open documents for material. Note that there is quite a large amount of documents on the net, so we can generate brand-new tests every time. One may say that adversaries can look for the phrases over the Internet and get several hints. Our system hides the sources by substituting the consonants of the phrases against such an attack.The mechanism is designed to imitate the phenomenon called “consonant gradation” of natural languages.

Text-mining on incident reports to find knowledge on industrial safety

To prevent accidents, it is very important to learn why and how past accidents occurred and escalated. The information of accidents is mostly recorded in natural language texts, which is not convenient to analyze the flow of events in the accidents. This paper proposes a method to recognize typical flow of events in a large set of text reports. By focusing two adjacent sentences, our system succeeded to detect typical pairs of predecessor word and successor word. Then we can recognize the typical flows of accidents.

Modeling the Relationship Between Stress and Appetite to Create a Dish Recommendation System Based on Desired Nutrients

Using decision trees, we investigated the relationship between stress condition and the nutrients contained in the consumed foods during the stress condition. The objective of this study is to develop a prospective dish recommendation system. Some explicit analyses of the nutrients preferred by people with high stress condition were performed. Thus, we were able to find the possibility of a dish recommending system based on nutrients and stress level.

Detection System for Distinguishing Between Initial Reading and Rereading of a Digital Document by Observing Focal Point Movement

This study moves toward improving digital textbooks by using line-of-sight to judge whether the reader is reading a text for the first time. We believe that this information can be used to provide more useful educational materials. For example, when reading a digital textbook a second (or subsequent) time, it might automatically summarize important points. Of the layouts examined, we found no gaze differences between the first and second readings with a screen vertically divided between text and figures. With a scrolling layout, there was a clearly observed difference between first and second readings, with shorter, fewer, and more narrowly focused gazes during the second reading.

Determination of the Optimal Timing for a Relaxing Stimulus

In this study, we investigated the relationship between the timing of a relaxation stimulus and its effect on a participant performing intellectual work. Experiments using light and sound stimuli showed that the relaxation effect is weakened or becomes stressful when the stimulus is provided during the work period, as opposed to the beginning. In addition, we found a possible synergistic effect could be achieved by combining a light and sound stimulus.