This thesis proposes novel algorithms for using Gaussian processes for Discriminative pose estimation. We overcome the traditional limitations of Gaussian processes, their cubic training complexity and their uni-modal predictive distribution by assembling them in a mixture of experts formulation. Our First contribution shows that by creating a large number of Fixed size Gaussian process experts, we can build a model that is able to scale to large data sets and accurately learn the multi-modal and non- linear mapping between image features and the subject’s pose. We demonstrate that this model gives state of the art performance compared to other discriminative pose estimation techniques.We then extend the model to automatically learn the size and location of each expert. Gaussian processes are able to accurately model non-linear functional regression problems where the output is given as a function of the input. However, when an individual Gaussian process is trained on data which contains multi-modalities, or varying levels of ambiguity, the Gaussian process is unable to accurately model the data. We propose a novel algorithm for learning the size and location of each expert in our mixture of Gaussian processes model to ensure that the training data of each expert matches the assumptions of a Gaussian process. We show that this model is able to out perform our previous mixture of Gaussian processes model.Our final contribution is a dynamics framework for inferring a smooth sequence of pose estimates from a sequence of independent predictive distributions. Discriminative pose estimation infers the pose of each frame independently, leading to jittery tracking results. Our novel algorithm uses a model of human dynamics to infer a smooth path through a sequence of Gaussian mixture models as given by our mixture of Gaussian processes model. We show that our algorithm is able to smooth and correct some mis- takes made by the appearance model alone, and outperform a baseline linear dynamical system.

No description available.

<p>Human action recognition continues to evolve and is examined better using deep learning techniques. Several successes have been recorded in the field of action recognition but only very few has focused on dance. This is because dance actions and, especially Traditional African dance, are long and involve fast movement of body parts. This research proposes a novel framework that applies data science algorithms to the field of cultural preservation by applying various deep learning techniques to identify, classify and model Traditional African dances from videos. Traditional African dances are important part of the African culture and heritage. Digital preservation of these dances in their myriad forms is a problem. The dance dataset was constituted using freely available YouTube videos. Three Traditional African dances – Adowa, Bata and Swange – were used for the dance classification process. Two Convolutional Neural Network (CNN) models were used for the classification and they achieved an accuracy of 97% and 98% respectively. Sound classification of Adowa, Bata and Swange drum ensembles were also carried out; an accuracy of 96% was achieved. Human Pose Estimation Algorithms were applied to the Sinte dance. A model of Sinte dance, which can be exported to other environments, was obtained.</p>

Pose estimation uses computer vision to predict how a body moves. The likeliness off different movements is predicted with a neural network and the most likely pose is predicted. With DeepLabCut, an open source software package for 3D animal pose estimation, information about animals behaviour and movement can be extracted. In this report the pose estimation of horses four hooves is used. By looking at the position of the hooves different gaits can be identified. Horses used for riding in the major disciplines in Sweden have three different gaits, walk, trot and canter. Walk is a four-stoke gait, trot is two-stoke and canter is three-stoke. This can be used to classify the different gaits. By looking at the hooves movement in vertical position over time and fitting a sinewave to the data it is possible to see the phase difference in the hooves movement. For walk and trot there was a significant pattern which was easy to identify and corresponded well to the theory of horses movement. For canter our pre-trained model lacked in accuracy, so the output data were insufficient. Therefore it was not possible to find a significant pattern for canter which corresponds to the theory of horses movements. The Fourier Transform were also tested to classify the gaits and when plotted it was possible to detect the different gaits, but not significant enough to be reliable for different horses in different sizes running in different paces. It was also possible to add the data for all four hooves together and fit a sinewave to the added data, and then compare it with the sinewaves for each hoof separately. Depending on the gait the frequency of the sinewaves differed between the hooves separately and added together and the gaits could be identified.

The task of 6D pose estimation with deep learning is to train networks to, from an im-age of an object, determine the rotation and translation of the object. Impressive resultshave recently been shown in deep learning based 6D pose estimation. However, many cur-rent solutions rely on real-world data when training, which as opposed to synthetic data,requires time consuming annotation. In this thesis, we introduce a pipeline for generatingsynthetic ground truth data for deep 6D pose estimation, where annotation is done auto-matically. With a 3D CAD-model, we use Blender to render 2D images of the model fromdifferent view points. We also create all other relevant data needed for pose estimation, e.g.,the poses of an object, mask images and 3D keypoints on the object. Using this pipeline, itis possible to adjust different settings to reduce the domain gap between synthetic data andreal-world data and get better pose estimation results. Such settings could be changing themethod of extracting 3D keypoints and varying the scale of the object or the light settingsin the scene.The network used to test the performance of training on our synthetic data is PVNet,which achieves state-of-the-art results for 6D pose estimation. This architecture learns tofind 2D keypoints of the object in the image, as well as 2D–3D keypoint correspondences.With these correspondences, the Perspective-n-Point (PnP) algorithm is used to extract apose. We evaluate the pose estimation of the different settings on the synthetic data andcompare these results to other state-of-the-art work. We find that using only real-worlddata for training is worse than using a combination of synthetic and real-world data. Sev-eral other findings are that varying scale and lightning, in addition to adding random back-ground images to the rendered images improves results. Four different novel keypoint se-lection methods are introduced in this work, and tried against methods used in previouswork. We observe that our methods achieve similar or better results. Finally, we use thebest possible settings from the synthetic data pipeline, but with memory limitations on theamount of training data. We are close to state-of-the-art results, and could get closer withmore data.

The proposed thesis of this dissertation has both a practical element and theoretical component which aim to answer key questions related to the use of angles-only navigation for autonomous orbital rendezvous. The first and fundamental principle to this work argues that an angles-only navigation filter can determine the relative position and orientation (pose) between two spacecraft to perform the necessary maneuvers and close proximity operations for autonomous orbital rendezvous. Second, the implementation of angles-only navigation for on-orbit applications is looked upon with skeptical eyes because of its perceived limitation of determining the relative range between two vehicles. This assumed, yet little understood subtlety can be formally characterized with a closed-form analytical observability criteria which specifies the necessary and sufficient conditions for determining the relative position and velocity with only angular measurements. With a mathematical expression of the observability criteria, it can be used to 1) identify the orbital rendezvous trajectories and maneuvers that ensure the relative position and velocity are observable for angles-only navigation, 2) quantify the degree or level of observability and 3) compute optimal maneuvers that maximize observability. In summary, the objective of this dissertation is to provide both a practical and theoretical foundation for the advancement of autonomous orbital rendezvous through the use of angles-only navigation.

In this thesis, a study of two blooming fields in the artificial intelligence topic is carried out. The first part of the present document is about 3D object recognition methods. Object recognition in general is about providing the ability to understand what objects appears in the input data of an intelligent system. Any robot, from industrial robots to social robots, could benefit of such capability to improve its performance and carry out high level tasks. In fact, this topic has been largely studied and some object recognition methods present in the state of the art outperform humans in terms of accuracy. Nonetheless, these methods are image-based, namely, they focus in recognizing visual features. This could be a problem in some contexts as there exist objects that look alike some other, different objects. For instance, a social robot that recognizes a face in a picture, or an intelligent car that recognizes a pedestrian in a billboard. A potential solution for this issue would be involving tridimensional data so that the systems would not focus on visual features but topological features. Thus, in this thesis, a study of 3D object recognition methods is carried out. The approaches proposed in this document, which take advantage of deep learning methods, take as an input point clouds and are able to provide the correct category. We evaluated the proposals with a range of public challenges, datasets and real life data with high success. The second part of the thesis is about hand pose estimation. This is also an interesting topic that focuses in providing the hand's kinematics. A range of systems, from human computer interaction and virtual reality to social robots could benefit of such capability. For instance to interface a computer and control it with seamless hand gestures or to interact with a social robot that is able to understand human non-verbal communication methods. Thus, in the present document, hand pose estimation approaches are proposed. It is worth noting that the proposals take as an input color images and are able to provide 2D and 3D hand pose in the image plane and euclidean coordinate frames. Specifically, the hand poses are encoded in a collection of points that represents the joints in a hand, so that they can be easily reconstructed in the full hand pose. The methods are evaluated on custom and public datasets, and integrated with a robotic hand teleoperation application with great success.

No description available.

In this thesis, a study of two blooming fields in the artificial intelligence topic is carried out. The first part of the present document is about 3D object recognition methods. Object recognition in general is about providing the ability to understand what objects appears in the input data of an intelligent system. Any robot, from industrial robots to social robots, could benefit of such capability to improve its performance and carry out high level tasks. In fact, this topic has been largely studied and some object recognition methods present in the state of the art outperform humans in terms of accuracy. Nonetheless, these methods are image-based, namely, they focus in recognizing visual features. This could be a problem in some contexts as there exist objects that look alike some other, different objects. For instance, a social robot that recognizes a face in a picture, or an intelligent car that recognizes a pedestrian in a billboard. A potential solution for this issue would be involving tridimensional data so that the systems would not focus on visual features but topological features. Thus, in this thesis, a study of 3D object recognition methods is carried out. The approaches proposed in this document, which take advantage of deep learning methods, take as an input point clouds and are able to provide the correct category. We evaluated the proposals with a range of public challenges, datasets and real life data with high success. The second part of the thesis is about hand pose estimation. This is also an interesting topic that focuses in providing the hand's kinematics. A range of systems, from human computer interaction and virtual reality to social robots could benefit of such capability. For instance to interface a computer and control it with seamless hand gestures or to interact with a social robot that is able to understand human non-verbal communication methods. Thus, in the present document, hand pose estimation approaches are proposed. It is worth noting that the proposals take as an input color images and are able to provide 2D and 3D hand pose in the image plane and euclidean coordinate frames. Specifically, the hand poses are encoded in a collection of points that represents the joints in a hand, so that they can be easily reconstructed in the full hand pose. The methods are evaluated on custom and public datasets, and integrated with a robotic hand teleoperation application with great success.

No description available.