Human pose estimation represents the skeleton of a person in color or depth images to improve a machine’s understanding of human movement. 3D human pose estimation uses a three-dimensional skeleton to represent the human body posture, which is more stereoscopic than a two-dimensional skeleton. Therefore, 3D human pose estimation can enable machines to play a role in physical education and health recovery, reducing labor costs and the risk of disease transmission. However, the existing datasets for 3D pose estimation do not involve fast motions that would cause optical blur for a monocular camera but would allow the subjects’ limbs to move in a more extensive range of angles. The existing models cannot guarantee both real-time performance and high accuracy, which are essential in physical education and health recovery applications. To improve real-time performance, researchers have tried to minimize the size of the model and have studied more efficient deployment methods. To improve accuracy, researchers have tried to use heat maps or point clouds to represent features, but this increases the difficulty of model deployment. To address the lack of datasets that include fast movements and easy-to-deploy models, we present a human kinetic dataset called the Kivi dataset and a hybrid model that combines the benefits of a heat map-based model and an end-to-end model for 3D human pose estimation. We describe the process of data collection and cleaning in this thesis. Our proposed Kivi dataset contains large-scale movements of humans. In the dataset, 18 joint points represent the human skeleton. We collected data from 12 people, and each person performed 38 sets of actions. Therefore, each frame of data has a corresponding person and action label. We design a preliminary model and propose an improved model to infer 3D human poses in real time. When validating our method on the Invariant Top-View (ITOP) dataset, we found that compared with the initial model, our improved model improves the mAP@10cm by 29%. When testing on the Kivi dataset, our improved model improves the mAP@10cm by 15.74% compared to the preliminary model. Our improved model can reach 65.89 frames per second (FPS) on the TensorRT platform.

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.

Understanding humans in visual content is fundamental for numerous computer vision applications. Extensive research has been conducted in the field of human pose estimation (HPE) to accurately locate joints and construct body representations from images and videos. Expanding on HPE, human mesh recovery (HMR) addresses the more complex task of estimating the 3D pose and shape of the entire human body. HPE and HMR have gained significant attention due to their applications in areas such as digital human avatar modeling, AI coaching, and virtual reality [135]. However, HPE and HMR come with notable challenges, including intricate body articulation, occlusion, depth ambiguity, and the limited availability of annotated 3D data. Despite the progress made so far, the research community continues to strive for robust, accurate, and efficient solutions in HPE and HMR, advancing us closer to the ultimate goals in the field. This dissertation tackles various challenges in the domains of HPE and HMR. The initial focus is on video-based HPE, where we proposed a transformer architecture named PoseFormer [136] to leverage to capture the spatial relationships between body joints and temporal correlations across frames. This approach effectively harnesses the comprehensive connectivity and expressive power of transformers, leading to improved pose estimation accuracy in video sequences. Building upon this, the dissertation addresses the heavy computational and memory burden associated with image-based HMR. Our proposed Feater Map-based Transformer method (FeatER [133]) and a Pooling attention transformer method (POTTER[130]), demonstrate superior performance while significantly reducing computational and memory requirements compared to existing state-of-the-art techniques. Furthermore, a diffusion-based framework (DiffMesh[134]) is proposed for reconstructing high-quality human mesh outputs given input video sequences. These achievements provide practical and efficient solutions that cater to the demands of real-world applications in HPE and HMR. In this dissertation, our contributions advance the fields of HPE and HMR, bringing us closer to accurate and efficient solutions for understanding humans in visual content.

Human pose estimation (HPE) is an ever-growing research field, with an increasing number of publications in the computer vision and deep learning fields and it covers a multitude of practical scenarios, from sports to entertainment and from surveillance to medical applications. Despite the impressive results that can be obtained with HPE, there are still many problems that need to be tackled when dealing with real-world applications. Most of the issues are linked to a poor or completely wrong detection of the pose that emerges from the inability of the network to model the viewpoint. This thesis shows how designing viewpoint-equivariant neural networks can lead to substantial improvements in the field of human pose estimation, both in terms of state-of-the-art results and better real-world applications. By jointly learning how to build hierarchical human body poses together with the observer viewpoint, a network can learn to generalise its predictions when dealing with previously unseen viewpoints. As a result, the amount of training data needed can be drastically reduced, simultaneously leading to faster and more efficient training and more robust and interpretable real-world applications.

This thesis presents new methods in two closely related areas of computer vision: human pose estimation, and gesture recognition in videos. In human pose estimation, we show that random forests can be used to estimate human pose in monocular videos. To this end, we propose a co-segmentation algorithm for segmenting humans out of videos, and an evaluator that predicts whether the estimated poses are correct or not. We further extend this pose estimator to new domains (with a transfer learning approach), and enhance its predictions by predicting the joint positions sequentially (rather than independently) in an image, and using temporal information in the videos (rather than predicting the poses from a single frame). Finally, we go beyond random forests, and show that convolutional neural networks can be used to estimate human pose even more accurately and efficiently. We propose two new convolutional neural network architectures, and show how optical flow can be employed in convolutional nets to further improve the predictions. In gesture recognition, we explore the idea of using weak supervision to learn gestures. We show that we can learn sign language automatically from signed TV broadcasts with subtitles by letting algorithms 'watch' the TV broadcasts and 'match' the signs with the subtitles. We further show that if even a small amount of strong supervision is available (as there is for sign language, in the form of sign language video dictionaries), this strong supervision can be combined with weak supervision to learn even better models.

Motion capture has in recent years grown in interest in many fields from both game industry to sport analysis. The need of reflective markers and expensive multi-camera systems limits the business since they are costly and time-consuming. One solution to this could be a deep neural network trained to extract 3D joint estimations from a 2D video captured with a smartphone. This master thesis project has investigated the accuracy of a trained convolutional neural network, MargiPose, that estimates 25 joint positions in 3D from a 2D video, against a gold standard, multi-camera Vicon-system. The project has also investigated if the data from the deep neural network can be connected to a biomechanical modelling software, AnyBody, for further analysis. The final intention of this project was to analyze how accurate such a combination could be in golf swing analysis. The accuracy of the deep neural network has been evaluated with three parameters: marker position, angular velocity and kinetic energy for different segments of the human body. MargiPose delivers results with high accuracy (Mean Per Joint Position Error (MPJPE) = 1.52 cm) for a simpler movement but for a more advanced motion such as a golf swing, MargiPose achieves less accuracy in marker distance (MPJPE = 3.47 cm). The mean difference in angular velocity shows that MargiPose has difficulties following segments that are occluded or has a greater motion, such as the wrists in a golf swing where they both move fast and are occluded by other body segments. The conclusion of this research is that it is possible to connect data from a trained CNN with a biomechanical modelling software. The accuracy of the network is highly dependent on the intention of the data. For the purpose of golf swing analysis, this could be a great and cost-effective solution which could enable motion analysis for professionals but also for interested beginners. MargiPose shows a high accuracy when evaluating simple movements. However, when using it with the intention of analyzing a golf swing in i biomechanical modelling software, the outcome might be beyond the bounds of reliable results.

My work focuses on taking a single person as input and predicting the intentional movement of one dance partner based on the other dance partner's movement. Human pose estimation has been applied to dance and computer vision, but many existing applications focus on a single individual or multiple individuals performing. Currently there are very few works that focus specifically on dance couples combined with pose prediction. This thesis is applicable to the entertainment and gaming industry by training people to dance with a virtual dance partner. Many existing interactive or virtual dance partners require a motion capture system, multiple cameras or a robot which creates an expensive cost. This thesis does not use a motion capture system and combines OpenPose with swing dance YouTube videos to create a virtual dance partner. By taking in the current dancer's moves as input, the system predicts the dance partner's corresponding moves in the video frames. In order to create a virtual dance partner, datasets that contain information about the skeleton keypoints are necessary to predict a dance partner's pose. There are existing dance datasets for a specific type of dance, but these datasets do not cover swing dance. Furthermore, the dance datasets that do include swing have a limited number of videos. The contribution of this thesis is a large swing dataset that contains three different types of swing dance: East Coast, Lindy Hop and West Coast. I also provide a basic framework to extend the work to create a real-time and interactive dance partner.

No description available.

This dissertation focuses on applying artificial intelligence techniques to human-robot interaction, which involves human pose estimation and dual-arm robotic manipulation. The motivating application behind this work is autonomous victim extraction in disaster scenarios using a conceptual design of a Semi-Autonomous Victim Extraction Robot (SAVER). SAVER is equipped with an advanced sensing system and two powerful robotic manipulators as well as a head and neck stabilization system to achieve autonomous safe and effective victim extraction, thereby reducing the potential risk to field medical providers. This dissertation formulates the autonomous victim extraction process using a dual-arm robotic manipulation system for human-robot interaction. According to the general process of Human-Robot Interaction (HRI), which includes perception, control, and decision-making, this research applies machine learning techniques to human pose estimation, robotic manipulator modeling, and dual-arm robotic manipulation, respectively. In the human pose estimation, an efficient parallel ensemble-based neural network is developed to provide real-time human pose estimation on 2D RGB images. A 13-limb, 14-joint skeleton model is used in this perception neural network and each ensemble of the neural network is designed for a specific limb detection. The parallel structure poses two main benefits: (1) parallel ensembles architecture and multiple Graphics Processing Units (GPU) make distributed computation possible, and (2) each individual ensemble can be deployed independently, making the processing more efficient when the detection of only some specific limbs is needed for the tasks. Precise robotic manipulator modeling benefits from the simplicity of the controller design and improves the performance of trajectory following. Traditional system modeling relies on first principles, simplifying assumptions and prior knowledge. Any imperfection in the above could lead to an analytical model that is different from the real system. Machine learning techniques have been applied in this field to pursue faster computation and more accurate estimation. However, a large dataset is always needed for these techniques, while obtaining the data from the real system could be costly in terms of both time and maintenance. In this research, a series of different Generative Adversarial Networks (GANs) are proposed to efficiently identify inverse kinematics and inverse dynamics of the robotic manipulators. One four-Degree-of-Freedom (DOF) robotic manipulator and one six-DOF robotic manipulator are used with different sizes of the dataset to evaluate the performance of the proposed GANs. The general methods can also be adapted to other systems, whose dataset is limited using general machine learning techniques. In dual-arm robotic manipulation, basic behaviors such as reaching, pushing objects, and picking objects up are learned using Reinforcement Learning. A Teacher-Student advising framework is proposed to learn a single neural network to control dual-arm robotic manipulators with previous knowledge of controlling a single robotic manipulator. Simulation and experimental results present the efficiency of the proposed framework compared to the learning process from scratch. Another concern in robotic manipulation is safety constraints. A variable-reward hierarchical reinforcement learning framework is proposed to solve sparse reward and tasks with constraints. A task of picking up and placing two objects to target positions while keeping them in a fixed distance within a threshold is used to evaluate the performance of the proposed method. Comparisons to other state-of-the-art methods are also presented. Finally, all the three proposed components are integrated as a single system. Experimental evaluation with a full-size manikin was performed to validate the concept of applying artificial intelligence techniques to autonomous victim extraction using a dual-arm robotic manipulation system. / Doctor of Philosophy / Using mobile robots for autonomous victim extraction in disaster scenarios reduces the potential risk to field medical providers. This dissertation focuses on applying artificial intelligence techniques to this human-robot interaction task involving pose estimation and dual-arm manipulation for victim extraction. This work is based on a design of a Semi-Autonomous Victim Extraction Robot (SAVER). SAVER is equipped with an advanced sensing system and two powerful robotic manipulators as well as a head and neck stabilization system attached on an embedded declining stretcher to achieve autonomous safe and effective victim extraction. Therefore, the overall research in this dissertation addresses: human pose estimation, robotic manipulator modeling, and dual-arm robotic manipulation for human pose adjustment. To accurately estimate the human pose for real-time applications, the dissertation proposes a neural network that could take advantages of multiple Graphics Processing Units (GPU). Considering the cost in data collection, the dissertation proposed novel machine learning techniques to obtain the inverse dynamic model and the inverse kinematic model of the robotic manipulators using limited collected data. Applying safety constraints is another requirement when robots interacts with humans. This dissertation proposes reinforcement learning techniques to efficiently train a dual-arm manipulation system not only to perform the basic behaviors, such as reaching, pushing objects and picking up and placing objects, but also to take safety constraints into consideration in performing tasks. Finally, the three components mentioned above are integrated together as a complete system. Experimental validation and results are discussed at the end of this dissertation.

Nonrigid deformation modeling and estimation from images is a technically challenging task due to its nonlinear, nonconvex and high-dimensional nature. Traditional optimization procedures often rely on good initializations and give locally optimal solutions. On the other hand, learning-based methods that directly model the relationship between deformed images and their parameters either cannot handle complicated forms of mapping, or suffer from the Nyquist Limit and the curse of dimensionality due to high degrees of freedom in the deformation space. In particular, to achieve a worst-case guarantee of ∈ error for a deformation with d degrees of freedom, the sample complexity required is O(1/∈d). In this thesis, a generative model for deformation is established and analyzed using a unified theoretical framework. Based on the framework, three algorithms, Data-Driven Descent, Top-down and Bottom-up Hierarchical Models, are designed and constructed to solve the generative model. Under Lipschitz conditions that rule out unsolvable cases (e.g., deformation of a blank image), all algorithms achieve globally optimal solutions to the specific generative model. The sample complexity of these methods is substantially lower than that of learning-based approaches, which are agnostic to deformation modeling. To achieve global optimality guarantees with lower sample complexity, the structureembedded in the deformation model is exploited. In particular, Data-driven Descentrelates two deformed images that are far away in the parameter space by compositionalstructures of deformation and reduce the sample complexity to O(Cd log 1/∈).Top-down Hierarchical Model factorizes the local deformation into patches once theglobal deformation has been estimated approximately and further reduce the samplecomplexity to O(Cd/1+C2 log 1/∈). Finally, the Bottom-up Hierarchical Model buildsrepresentations that are invariant to local deformation. With the representations, theglobal deformation can be estimated independently of local deformation, reducingthe sample complexity to O((C/∈)d0) (d0 ≪ d). From the analysis, this thesis showsthe connections between approaches that are traditionally considered to be of verydifferent nature. New theoretical conjectures on approaches like Deep Learning, arealso provided. practice, broad applications of the proposed approaches have also been demonstrated to estimate water distortion, air turbulence, cloth deformation and human pose with state-of-the-art results. Some approaches even achieve near real-time performance. Finally, application-dependent physics-based models are built with good performance in document rectification and scene depth recovery in turbulent media.