Particular motions are important to play sports with high performance. The particular motions are mastered by learning motions, and visual information is considered to be effective for understanding and learning motions. In recent years, HMD with VR has been introduced as a new tool for learning motions with visual information. An advantage of the HMD-based motion learning method is that it enables learners to switch their observation view. Here, this research investigates basic view characteristics of observing and reproducing particular dynamic motions, which would be necessary to develop some methods for switching observation view properly. An experiment was conducted in order to study the basic view characteristics. As for the observation view factor, we prepared two factor levels, one was the front mirror view, and the other the rear camera view. In the experiment, a subject recognized and reproduced some reference dynamic motions on real time with each of the two views. The experimental results revealed that the reproduction performance with the rear camera view was significantly better than that with the front mirror view in the case of the depth-directional motions, compared with the other case of the depth-uncorrelated motions. It should be noted that the difference in the motion reproduction may become crucial for learners in particular as the motion velocity increases. It is supposed that the observation with the front mirror view requires some mental transformation operation when the learners reproduce motions. In selecting the observation view, it is required to minimize the mental transformation operation. The requirement is expected to be satisfied with the rear camera view, provided that occlusions are not crucial for learners to observe reference motions.

Remote teleoperation of robotic manipulators requires a robust machine vision system in order to perform accurate movements in the navigated environment. Even though a 3D CAD model is available, the dimensions and poses of its components are subject to change due to extreme conditions. Integration of a stereoscopic camera into the control chain enables more precise object detection, pose-estimation, and tracking. However, the conventional stereoscopic pose-estimation methods still lack robustness and accuracy in the presence of harsh environmental conditions, such as high levels of radiation, deficient illumination, shiny metallic surfaces, etc. In this paper we investigate the ability of a specifically tuned iterative closest point (ICP) algorithm to operate in the aforementioned environments and suggest algorithmic improvements. We demonstrate that the proposed algorithm outperforms current state-of-the-art methods in both robustness and accuracy. The experiments are performed with a real robotic manipulator prototype and a stereoscopic machine vision system.

A flashover occurs when a fire spreads very rapidly through crevices due to intense heat. Flashovers present one of the most frightening and challenging fire phenomena to those who regularly encounter them: firefighters. Firefighters' safety and lives often depend on their ability to predict flashovers before they occur. Typical pre-flashover fire characteristics include dark smoke, high heat, and rollover ("angel fingers") and can be quantified by color, size, and shape. Using a color video stream from a firefighter's body camera, we applied generative adversarial neural networks for image enhancement. The neural networks were trained to enhance very dark fire and smoke patterns in videos and monitor dynamic changes in smoke and fire areas. Preliminary tests with limited flashover training videos showed that we predicted a flashover as early as 55 seconds before it occurred.

The rapid development of Unmanned Aerial Vehicle (UAV) technology, -also known as drones- has raised concerns on the safety of critical locations such as governmental buildings, nuclear stations, crowded places etc. Computer vision based approach for detecting these threats seems as a viable solution due to various advantages. We envision an autonomous drone detection and tracking system for the protection of strategic locations. It has been reported numerous times that, one of the main challenges for aerial object recognition with computer vision is discriminating birds from the targets. In this work, we have used 2-dimensional scale, rotation and translation invariant Generic Fourier Descriptor (GFD) features and classified targets as a drone or bird by a neural network. For the training of this system, a large dataset composed of birds and drones is gathered from open sources. We have achieved up to 85.3% overall correct classification rate.

Traditional quality estimators evaluate an image's resemblance to a reference image. However, quality estimators are not well suited to the similar but somewhat different task of utility estimation, where an image is judged instead by how useful it would be in terms of extracting information about image content. While full-reference image utility metrics have been developed which outperform quality estimators for the utility-prediction task, assuming the existence of a high-quality reference image is not always realistic. The Oxford Visual Geometry Group's (VGG) deep convolutional neural network (CNN) [1], designed for object recognition, is modified and adapted to the task of utility estimation. This network achieves no-reference utility estimation performance near the full-reference state of the art, with a Pearson correlation of 0.946 with subjective utility scores of the CU-Nantes database and root mean square error of 12.3. Other noreference techniques adapted from the quality domain yield inferior performance. The CNN also generalizes better to distortion types outside of the training set, and is easily updated to include new types of distortion. Early stages of the network apply transformations similar to those of previously developed full-reference utility estimation algorithms.

This paper proposes a guitar fingering assessing system based on CNN (Convolutional Neural Network) hand pose estimation and SVR (Support Vector Regression) evaluation. To spur our progress, first, a CNN architecture is proposed to estimate temporal 3D position of 16 joints of hand; then, based on a DCT (Discrete Cosine Transform) feature and SVR, fingering of guitarist is scored to interpret how well guitarist played. We also release a new dataset for professional guitar playing analysis with significant advantage in total number of video, professional judgement by expert of guitarist, accurate annotation for hand pose and score of guitar performance. Experiments using videos containing multiple persons' guitar plays under different conditions demonstrate that the proposed method outperforms the current state-of-art with (1) low mean error (Euclid distance of 6,1 mm) and high computation efficiency for hand pose estimation; (2) high rank correlation (0.68) for assessing the fingering (C major scale and symmetrical excise) of guitarists.

To contribute to the actualization of the care worker assistance robot, this paper proposes a method for detecting whether the care receiver is chewing from the video sequence acquired by the camera that observes that receiver. The proposed method detects the receiver's face and areas for both cheeks and chin. After applying some normalization to the areas, chewing detection that uses a variable-intensity template is performed, where the template consists of shape models, interest points and intensity distribution model. A likelihood based on the variable-intensity template is computed so that the receiver is judged whether the receiver is chewing. Experiments using seven subjects are conducted. As a result, the accuracy of chewing detection by the proposed method is 83%, which is quite promising.

Towards the establishment of the preventive medical care for the cerebral aneurysm, this paper proposes an SVM based method for building a discrimination function that classifies the presence or absence of the cerebral aneurysm using the cerebral blood vessel's shape features obtained from medical images such as MR images. Using the discrimination function, this paper explores how much each feature affects the onset of the cerebral aneurysm. This paper deals with the internal carotid artery (ICA). The blood vessel (ICA)'s shape features are extracted from medical images of 18 persons without cerebral aneurysm and 13 patients with a cerebral aneurysm. From the medical image, the cross sections and centerline of the ICA are obtained. The cross sections are divided into nine sections along the centerline. Shape features such as the cross sectional area, its circularity, curvature, torsion, length of the centerline and branch angles are obtained in each section; as a total, 113 features including the mean and variance of some features in each section are used for building the SVM. As a result of conducting the experiments, the accuracy for discriminating the presence/absence of the aneurysm by the SVM is 90.3%. In the obtained discrimination function, the coefficient values of the function can be considered how much the features affect the onset of the aneurysm. The features that could significantly cause the onset of the cerebral aneurysm are clarified, and the reasons why these features are significant are discussed.

Outlier detection (OD) has been popularly developed in many fields such as medical diagnosis, network intrusion detection, fraud detection and military surveillance. This paper presents an accumulated relative density (ARD) OD method to identify outliers which possess relatively low or high local density. Previously, many density-based OD methods, such as local outlier factor (LOF) and Local Correlation Integral (LOCI), are applied to detect outliers which have low relative density in the data set. Relative local density (RLD) is measured and then compared with each other by statistics to label abnormities. In the proposed ARD method, a big circle centered at every data point is formed first. This big circle covers some data points with its radius. Then, for each encapsulated point inside this big circle, a small circle centered at itself is defined. Afterward, the ratio of number of covered data points inside the small circle of that particular point to the average number of data points in all small circles is defined as the RLD. After RLDs of all data points are calculated, a point whose RLD deviates greatly from the mean of all RLDs will be labeled as an outlier, otherwise as inliers. This ARD method was evaluated by a real world traffic data set which was originally represented as spatial-temporal (ST) traffic flow signals. The ST signals were processed by a principal component analysis (PCA) to reduce its dimension into two-dimensional 2D data points. An average 95% detection success rate (DSR) of OD can be achieved by this method.