The ultra-wideband (UWB) technology has a vast unlicensed frequency spectrum, which can support precise indoor positioning in orders of centimeters. The features of UWB signals can be utilized for variety of applications. In this project first we present an empirical channel models to analyze the localization accuracy of the UWB technology for interactive electronic gaming (Ping-Pong) in Line-of-Sight (LOS) and Obstructed LOS (OLOS) scenarios. Then we introduce a new concept that we refer to as micro-gesture detection to handle the more refined motions of the hand, such as rotation, while one antenna is held by the user using features of UWB signal. We use four specific features of the UWB signals: time of arrival, power of the first peak, total power, and the Root-Mean Square (RMS) of the delay spread, for this purpose. As the hand rotates the position of the antenna in the hand and the external antenna changes from LOS to OLOS. We demonstrate that features of the UWB signals are more useful than the RSS signal of the Wi-Fi to detect this class of micro-gestures. We foresee this micro-gesture detection capabilities become helpful for the people with limited ability or visually impaired for implementation of simplified sign language to communication with electronic devices located away from a person. We compare gesture detection using multiple features of the UWB signal with traditional gesture detection using the received signal strength (RSS) of the Wi-Fi signal.

This diploma thesis deals with the hand gesture detector. Theoretical part is devoted to the discussion of current segmentation and detection methods. They were chosen those, which would work best as the controlling methods of multimedia player. The proposed algorithm is described with the help of flowcharts. Detector is written in Java language with the use of OpenCV library. Background subtraction segmentation method and convex hull approximation are used for gesture detection. Detector can distinguish between six gestures and track the center of palm. Finally, the detector was connected with proposed and implemented document, image and multimeda files browser.

Des efforts importants ont été faits pour améliorer la précision de la détection des actions humaines à l’aide des articulations du squelette. Déterminer les actions dans un environnement bruyant reste une tâche difficile, car les coordonnées cartésiennes des articulations du squelette fournies par la caméra de détection à profondeur dépendent de la position de la caméra et de la position du squelette. Dans certaines applications d’interaction homme-machine, la position du squelette et la position de la caméra ne cessent de changer. La méthode proposée recommande d’utiliser des valeurs de position relatives plutôt que des valeurs de coordonnées cartésiennes réelles. Les récents progrès des réseaux de neurones à convolution (RNC) nous aident à obtenir une plus grande précision de prédiction en utilisant des entrées sous forme d’images. Pour représenter les articulations du squelette sous forme d’image, nous devons représenter les informations du squelette sous forme de matrice avec une hauteur et une largeur égale. Le nombre d’articulations du squelette fournit par certaines caméras de détection à profondeur est limité, et nous devons dépendre des valeurs de position relatives pour avoir une représentation matricielle des articulations du squelette. Avec la nouvelle représentation des articulations du squelette et le jeu de données MSR, nous pouvons obtenir des performances semblables à celles de l’état de l’art. Nous avons utilisé le décalage d’image au lieu de l’interpolation entre les images, ce qui nous aide également à obtenir des performances similaires à celle de l’état de l’art. / There have been significant efforts in the direction of improving accuracy in detecting human action using skeleton joints. Recognizing human activities in a noisy environment is still challenging since the cartesian coordinate of the skeleton joints provided by depth camera depends on camera position and skeleton position. In a few of the human-computer interaction applications, skeleton position, and camera position keep changing. The proposed method recommends using relative positional values instead of actual cartesian coordinate values. Recent advancements in CNN help us to achieve higher prediction accuracy using input in image format. To represent skeleton joints in image format, we need to represent skeleton information in matrix form with equal height and width. With some depth cameras, the number of skeleton joints provided is limited, and we need to depend on relative positional values to have a matrix representation of skeleton joints. We can show the state-of-the-art prediction accuracy on MSR data with the help of the new representation of skeleton joints. We have used frames shifting instead of interpolation between frames, which helps us achieve state-of-the-art performance.