We present a novel, real-time, markerless vision-based tracking system, employing a rigid orthogonal configuration of two pairs of opposing cameras. Our system uses optical flow over sparse features to overcome the limitation of vision-based systems that require markers or a pre-loaded model of the physical environment. We show how opposing cameras enable cancellation of common components of optical flow leading to an efficient tracking algorithm that captures five degrees of freedom including direction of translation and angular velocity. Experiments comparing our device with an electromagnetic tracker show that its average tracking accuracy is 80% over 185 frames, and it is able to track large range motions even in outdoor settings. We also present how opposing cameras in vision-based inside-looking-out systems can be used for gesture recognition. To demonstrate our approach, we discuss three different algorithms for recovering motion parameters at different levels of complete recovery. We show how optical flow in opposing cameras can be used to recover motion parameters of the multi-camera rig. Experimental results show gesture recognition accuracy of 88.0%, 90.7% and 86.7% for our three techniques, respectively, across a set of 15 gestures.

La thèse porte sur la détection d’objets à partir d’une caméra embarquée sur un véhicule mobile en exploitant l’approche monoculaire « c-vélocité ». Cette méthode s’inspire de la méthode appelée « v-disparité » utilisée en stéréovision : toutes deux ont pour objectif la détection d’objets en les approximant par des plans d’orientations différentes, ce qui permet d’éviter, en monoculaire, d’estimer la profondeur. Ces deux approches, monoculaires et binoculaires, permettent de transformer le problème complexe de la détection d’objets en un problème plus simple de détection de formes paramétriques simples (droites, paraboles) dans un nouvel espace de représentation où la détection peut être réalisée à l’aide d’une transformée de Hough. La « c-vélocité », pour être efficace, requiert un calcul assez précis du flot optique et une bonne estimation de la position du Foyer d’expansion (FOE). Dans cette thèse, nous avons étudié les approches existantes de calcul de flot optique et sommes arrivés à la conclusion qu’aucune n’est vraiment performante notamment sur les régions homogènes telle que la route dans les scènes qui correspondent à l’application que nous visons à savoir : les véhicules intelligents. Par ailleurs, les méthodes d’estimation du flot optique peinent également à fournir une bonne estimation dans le cas de déplacement importants dans les régions proches de la caméra. Nous proposons dans cette thèse d’exploiter à la fois un modèle 3D de la scène et une estimation approximative de la vitesse du véhicule à partir d’autres capteurs intégrés. L’utilisation de connaissances a priori permet de compenser le flot dominant pour faciliter l’estimation de la partie résiduelle par une approche classique. Par ailleurs, trois approches différentes sont proposées pour détecter le foyer d’expansion. Parmi elles, nous proposons une méthode novatrice permettant d’estimer le FOE en exploitant la norme du flot et la structure de la scène à partir d’un processus « c-vélocité » inversé. En plus d’améliorer ces étapes préliminaires, nous proposons aussi l’optimisation et l’accélération de l’algorithme « c-vélocité » par une implémentation multithread. Enfin, nous proposons une modification de l’approche c-vélocité d’origine afin d’anticiper une éventuelle coopération mouvement/stéréo, proposée en perspective, à travers un jumelage avec la v-disparité. / This thesis is based on the detection of objects from an onboard moving camera by exploiting the monocular approach "c-velocity". This method is inspired by the method called "v-disparity" used in stereovision: both methods aim at detecting objects by approximating objects into plans with different orientations. Such approximation can avoid to estimate the depth in monocularvision. These two approaches, monocular and binocular, allow to transform the complex objet détection problem into a more simple parametric forms (eg. lines) detection in a new space, where these formes can be easily extracted using Hough Transform.The “c-velocity”, to make it effective, requires an accurate computation of optical flow and a good estimation of the focus of expansion (FOE) location. Therefore, we have studied the existing approaches of optical flow estimation and arrived at the conclusion that none of them is really powerful especially on the homogeneous regions such as road surface. In addition, the optical flow estimation methods also struggle to provide a good estimate in the case of huge displacement in the areas close to the camera. We propose in this thesis to exploit both a 3D model of the scene and a rough estimate about the vehicle speed from other integrated sensors. Using a priori knowledge allows to compensate the dominant optical flow and to facilitate the estimation of the rest part by a classical approach. In addition, three different approaches are proposed to detect the focus of expansion. Among them, we propose a novel method for estimating FOE by leveraging the flow norm and the scene structure from an inverse “c-velocity“ process. In addition to improve these preliminary steps, we also propose an acceleration and optimization of the “c-velocity“ algorithm by a multi-thread implementation. Finally, we propose a modification to the original “c-velocity“ approach in order to anticipate a possible cooperation motion/stereo, proposed in perspective, with the “v-disparity“ approach.

As electronics become smaller and more capable, it has become possible to conduct meaningful and sophisticated satellite missions in a small form factor. However, the capability of small satellites and the range of possible applications are limited by the capabilities of several technologies, including attitude determination and control systems. This dissertation evaluates the use of image-based visual attitude propagation as a compliment or alternative to other attitude determination technologies that are suitable for miniature satellites. The concept lies in using miniature cameras to track image features across frames and extracting the underlying rotation. The problem of visual attitude propagation as a small satellite attitude determination system is addressed from several aspects: related work, algorithm design, hardware and performance evaluation, possible applications, and on-orbit experimentation. These areas of consideration reflect the organization of this dissertation. A “stellar gyroscope” is developed, which is a visual star-based attitude propagator that uses relative motion of stars in an imager’s field of view to infer the attitude changes. The device generates spacecraft relative attitude estimates in three degrees of freedom. Algorithms to perform the star detection, correspondence, and attitude propagation are presented. The Random Sample Consensus (RANSAC) approach is applied to the correspondence problem to successfully pair stars across frames while mitigating false-positive and false-negative star detections. This approach provides tolerance to the noise levels expected in using miniature optics and no baffling, and the noise caused by radiation dose on orbit. The hardware design and algorithms are validated using test images of the night sky. The application of the stellar gyroscope as part of a CubeSat attitude determination and control system is described. The stellar gyroscope is used to augment a MEMS gyroscope attitude propagation algorithm to minimize drift in the absence of an absolute attitude sensor. The stellar gyroscope is a technology demonstration experiment on KySat-2, a 1-Unit CubeSat being developed in Kentucky that is in line to launch with the NASA ELaNa CubeSat Launch Initiative. It has also been adopted by industry as a sensor for CubeSat Attitude Determination and Control Systems (ADCS).