This thesis introduces a 3D body tracking system based on neutral networks and 3D geometry, which can robustly estimate body poses and accurate body joints. This system takes RGB-D data as input. Body poses and joints are firstly extracted from color image using deep learning approach. The estimated joints and skeletons are further translated to 3D space by using camera calibration information. This system is running at the rate of 3 4 frames per second. It can be used to any RGB-D sensors, such as Kinect, Intel RealSense [14] or any customized system with color depth calibrated. Comparing to the sate-of-art 3D body tracking system, this system is more robust, and can get much more accurate joints locations, which will benefits projects require precise joints, such as virtual try-on, body measure, real-time avatar driven.

Background: In humans, total body water volume and osmolality are tightly regulated by various homeostatic mechanisms, triggered by deviations in osmolality. Heat and exercise are two stressors, which in combination can cause dehydration, and an increase in fluid osmolality, contributing to health detriments, as well as deficits in aerobic exercise performance. However, it is unclear whether dehydration affects muscular strength. Deep-ocean mineral water has been shown to have benefits on various physiological and pathophysiological conditions, including aerobic performance and muscle strength. Objectives: The aims of this study were to examine any sex differences that may exist in response to dehydration of 3% of body mass, rehydration with various fluids, and the consequences of dehydration and rehydration on muscle power and hydration status. Design: We used a counterbalanced, crossover study design, in which subjects (n=17, 9 males vs 8 females) performed a dehydrating exercise protocol until achieving a 3% body mass loss, and then rehydrated with either deep-ocean mineral water (Deep), mountain spring water (Spring), or a carbohydrate-based sports drink (Sports). Subjects completed the protocol three times, with each subject receiving the rehydrating fluid in a different order to control for order effects. Saliva samples were collected throughout the protocol to measure osmolality, and muscle strength was measured by peak torque leg extension at baseline, post-exercise, and post-rehydration. Results: We found no differences between men and women in baseline or peak salivary osmolality, or in the exercise-induced increase in osmolality. Male subjects took less time to reach 3% body mass loss than females, and females demonstrated lower sweat rates than males. Salivary osmolality returned to baseline after rehydration, with the Deep group exhibiting a significantly more rapid return to baseline, for both sexes, compared to Sports and Spring. Males generated greater peak torque extension than females at baseline, while both males and females displayed a similar significant deficit in this measure following dehydration. Peak torque recovery post-rehydration was significantly affected by fluid designation and sex, and a significant difference was seen between the Deep and the Sports groups in females. Conclusions: Males reached 3% body mass loss faster than females, while dehydration resulted in increased salivary osmolality and muscle strength deficits similarly for males and females. Deep-ocean mineral water had a significant beneficial effect on hydration recovery, for both males and females, compared to the other fluids. Recovery of muscle strength after rehydration was affected by fluid and sex, with the main driver being females.

This study investigates the impact of employing the Miller cycle on a high specific power downsized gasoline engine by means of Early Intake Valve Closing (EIVC) and Late Intake Valve Closing (LIVC). This investigation assesses the potential for the Miller cycle to improve fuel economy at part load points, as well as high load points with significantly elevated boost pressures (Deep Miller) of up to 4 bar abs. The impact of geometric Compression Ratio (CR) and Exhaust Back Pressure (EBP) has also been investigated. The knock mitigating qualities of Deep Miller have been assessed, and its ability to increase maximum engine load explored. Low Speed Pre-ignition (LSPI) and autoignition tendencies with reduced coolant flow rates and with aged and new fuels have also been studied. This study comprises both experimental and analytical studies. A Ricardo Hydra single cylinder thermodynamic engine was developed and used for the experimental component of the study. This engine features a high specific power output (120kW/l) cylinder head from the Mahle 1.2l 3 cylinder aggressively downsized engine. The analytical component was carried out using a 1-dimensional GT-Power model based on the Ricardo Hydra experimental engine. A Design of Experiments (DoE) based test plan was adopted for this analytical study. The experimental study found that EIVC was the optimal strategy for improving fuel economy at both part-load and high-load conditions. LIVC yielded a fuel economy penalty at part-load operations and a fuel economy improvement at high-loads. The unexpected part-load LIVC result was attributed to the engine breathing dynamics of the experimental engine. The analytical study found moderate LIVC to be the optimal strategy at lower speeds, unless compensation for the increased degree of scavenging experienced with EIVC was compensated for, in which case EIVC was optimum. At higher speeds EIVC was found to be optimum regardless of whether or not compensation for scavenging was employed. It was generally found that less sensitivity to EBP was exhibited the more extreme the EIVC and LIVC. It was also found that a higher geometric CR could be tolerated with extreme EIVC and LIVC, and a fuel economy benefit could be obtained through the elevation of Geometric CR.

This thesis tackles the problem of large-scale visual search for categories within large collections of images. Given a textual description of a visual category, such as 'car' or 'person', the objective is to retrieve images containing that category from the corpus quickly and accurately, and without the need for auxiliary meta-data or, crucially and in contrast to previous approaches, expensive pre-training. The general approach to identifying different visual categories within a dataset is to train classifiers over features extracted from a set of training images. The performance of such classifiers relies heavily on sufficiently discriminative image representations, and many methods have been proposed which involve the aggregating of local appearance features into rich bag-of-words encodings. We begin by conducting a comprehensive evaluation of the latest such encodings, identifying best-of-breed practices for training powerful visual models using these representations. We also contrast these methods with the latest breed of Convolutional Network (ConvNet) based features, thus developing a state-of-the-art architecture for large-scale image classification. Following this, we explore how a standard classification pipeline can be adapted for use in a real-time setting. One of the major issues, particularly with bag-of-words based methods, is the high dimensionality of the encodings, which causes ranking over large datasets to be prohibitively expensive. We therefore assess different methods for compressing such features, and further propose a novel cascade approach to ranking which both reduces ranking time and improves retrieval performance. Finally, we explore the problem of training visual models on-the-fly, making use of visual data dynamically collected from the web to train classifiers on demand. On this basis, we develop a novel GPU architecture for on-the-fly visual category search which is capable of retrieving previously unknown categories over unannonated datasets of millions of images in just a few seconds.

Magister Scientiae Dentium - MSc(Dent) / Purpose: To assess the value of postoperative CT scans in the management of severe cervicofacial infections or deep neck abscesses. Patients and Methods: Thirteen patients underwent post-operative contrast enhanced CT scans after initial incision and drainage followed by a tracheostomy. As per surgical protocol, the CT scans were taken with radiopaque surgical drains in situ from the brain down to the mediastinum and chest/lungs. Data were collected on the presence of abscesses, their location, the location of surgical drains and presence of other pathology. Results: The most common initially affected space was the submandibular space (69%), followed by the submental space (62%). Almost half of the cases studied had two spaces affected and slightly less than a third had three spaces affected. Slightly more than half (54%) of the patients were in good health generally. More than a third (38.5%) of the patients were HIV positive, and thus immuno-compromised. The mean time lapse between the CT and surgery was 2.61 days (SD = 1.56). In the majority (69.23%) of post-operative scans a residual pus collection was found. In just over a quarter (30.8 %) of the patients no residual pus collection was detected on the postoperative CT scan, while in seven patients (53.8%) affected fascial spaces were missed by the surgeon. These spaces included the submasseteric, anteriormediastinum and parotid spaces. Almost half (44.15%) of the patients required a repeatincision and drainage. Conclusion: This study demonstrates clinical value of post-operative CT scans in severe cervicofacial infections. This study also advances the argument for acquisition of preoperative CECT scans for all patients with severe cervicofacial infection.

Master of Science / Department of Computing and Information Sciences / William H. Hsu / The goal of this research is to present a comparison between different clustering and segmentation techniques, both supervised and unsupervised, to detect plant and crop rows. Aerial images, taken by an Unmanned Aerial Vehicle (UAV), of a corn field at various stages of growth were acquired in RGB format through the Agronomy Department at the Kansas State University. Several segmentation and clustering approaches were applied to these images, namely K-Means clustering, Excessive Green (ExG) Index algorithm, Support Vector Machines (SVM), Gaussian Mixture Models (GMM), and a deep learning approach based on Fully Convolutional Networks (FCN), to detect the plants present in the images. A Hough Transform (HT) approach was used to detect the orientation of the crop rows and rotate the images so that the rows became parallel to the x-axis. The result of applying different segmentation methods to the images was then used in estimating the location of crop rows in the images by using a template creation method based on Green Pixel Accumulation (GPA) that calculates the intensity profile of green pixels present in the images. Connected component analysis was then applied to find the centroids of the detected plants. Each centroid was associated with a crop row, and centroids lying outside the row templates were discarded as being weeds. A comparison between the various segmentation algorithms based on the Dice similarity index and average run-times is presented at the end of the work.

This Thesis explores the connectiveness of the West Florida Shelf’s various areas of economic and ecological importance by considering five case studies of varying dynamic forcing influences and time. The advection of water about the shelf moves nutrients and has a direct impact on the shelf’s ecology and the determination of whether or not the shelf will be oligotrophic at any given time or location. The case studies are analyzed both quantitatively and qualitatively after quasi-isopycnal particle trajectory calculations are completed for each. The findings support a combination of local and deep-ocean forcing being ideal for the maximum advection and opportunity for potential connectivity between areas of the shelf, and provide a solid guide for moving forward with a considerable ensemble of studies in the future to approach the question from a statistical perspective. The numerical scheme used to calculate the particle trajectories is a 4th order Runge-Kutta method. The scheme is investigated for it appropriateness and pitfalls as a backward trajectory calculation tool by direct comparison between forward trajectory calculations and attempting to replicate the result in the backward direction. The findings support that the more linear the trajectory and the more restrictive the dynamics acting upon a particle at any given location, the better the backward and forward replication will be, although it is still an approximation, much like any other iterative tool used for approximating a solution to an ordinary differential equation.

Feature engineering is a process that augments the feature vector of a machine learning model with calculated values that are designed to enhance the accuracy of a model’s predictions. Research has shown that the accuracy of models such as deep neural networks, support vector machines, and tree/forest-based algorithms sometimes benefit from feature engineering. Expressions that combine one or more of the original features usually create these engineered features. The choice of the exact structure of an engineered feature is dependent on the type of machine learning model in use. Previous research demonstrated that various model families benefit from different types of engineered feature. Random forests, gradient-boosting machines, or other tree-based models might not see the same accuracy gain that an engineered feature allowed neural networks, generalized linear models, or other dot-product based models to achieve on the same data set. This dissertation presents a genetic programming-based algorithm that automatically engineers features that increase the accuracy of deep neural networks for some data sets. For a genetic programming algorithm to be effective, it must prioritize the search space and efficiently evaluate what it finds. This dissertation algorithm faced a potential search space composed of all possible mathematical combinations of the original feature vector. Five experiments were designed to guide the search process to efficiently evolve good engineered features. The result of this dissertation is an automated feature engineering (AFE) algorithm that is computationally efficient, even though a neural network is used to evaluate each candidate feature. This approach gave the algorithm a greater opportunity to specifically target deep neural networks in its search for engineered features that improve accuracy. Finally, a sixth experiment empirically demonstrated the degree to which this algorithm improved the accuracy of neural networks on data sets augmented by the algorithm’s engineered features.

This thesis describes the successful development and initial evaluation of a proof-of-concept wireless monitoring system for improving the effectiveness and safety of pneumatic compression therapy to help prevent deep vein thrombosis (DVT). In the development, an important objective was to make feasible the practical and commercial deployment of such improved therapy systems in future, by focusing on a cost-effective design and implementation. Over the years, pneumatic compression has been shown to be an effective solution for the prevention of DVT. However, different problems and complications related to the use of commercial pneumatic compression de-vices that typically include automatic pressure controllers and pneumatic compression sleeves have been reported. For example, one study reported a high percentage of improperly applied or nonfunctional pneumatic compression devices in routine usage. Technical problems, non-compliance, and human error were identified as the causes behind the failed therapies. Also, it was reported that dedicated in-service instruction did not improve the proper use of the pneumatic compression controllers and sleeves. In another study, significant unanticipated variations between expected and delivered pneumatic compression therapy were reported: expected therapy delivered only an average of 77.8% of the time during the therapy, and much of the time key values related to the outcome of the therapy were found to have variations great than 10%. Specific hazards have also been reported. For example, one patient developed acute compartment syndrome after wearing a pair of pneumatic compression sleeves with faulty pressure release valves. In another case, epidural analgesia masked a malfunction resulting from a reversed connection between four-way plastic tubing of the sleeves and the controller, exposing a patient to a hazardous pressure of around 300mmHg,blocking all blood flow for a prolonged period of time. Newer models of pneumatic compression sleeves and controllers from various manufacturers claim to improve therapy by, for example, increasing the peak blood flow velocity. However, there is no evidence in the published literature to support such claims. A published review of the literature from1970-2002 reached the conclusion that the most important factors in im-proving therapy with pneumatic compression devices, particularly during and after surgery, were the degree of conformance of delivered therapy to the prescribed therapy, patient compliance, and the appropriateness of the site of compression. The inability to monitor delivered therapy and patient compliance remains a problem in efforts to improve pneumatic compression therapy. The above-described problems were addressed in the successful development of the innovative prototype described in this thesis. This wireless monitoring system should improve the effectiveness and safety of pneumatic compression therapy. Also, innovative aspects of the system design allow for cost-effective integration into existing commercial controllers and sleeves. For example, an innovative and potentially patentable usage and reprocess indicator was developed for pneumatic compression sleeves to significantly improve their safety and to reduce their cost of use per patient. / Applied Science, Faculty of / Electrical and Computer Engineering, Department of / Graduate

In view of the recent catastrophes associated with deep excavations, there is an urgent need to provide vital guidelines on the design of the construction process. To develop a simple tool for predicting ground deformation around a deep excavation construction for preliminary design and decision-making purposes, small scale centrifuge models were made to observe the complicated mechanisms involved. A newly developed actuation system, with which the construction sequences ofpropping could be implemented, was developed, the new procedures were proven to give more realistic initial ground conditions before excavation with minimal development of pre-excavation bending moment and wall displacement. Incremental wall deformation profiles generally followed the O'Rourke cosine bulge equation and a new deformation mechanism was proposed with respect to wall toe fixity and excavation geometry. Validation of the conservation energy principle was carried out for the undrained excavation process. The total loss of potential energy was shown to be balanced by the total work done in shearing and the total elastic energy stored in structures with an error term of 30%. An improved mobilizable strength method (MSD) method using observed mechanistic deformation patterns was introduced to calculate the displacement profile of a multi-propped undrained excavation in soft clay. The incremental loss in potential energy associated with the formation of settlement toughs was balanced by the sum of incremental storage of elastic energy and the energy dissipation in shearing. A reasonable agreement was found between the prediction by the MSD method and the finite element results computed by an advanced MIT-E3 model for wall displacements, ground settlement, base heave and bending moment on fixed base walls. For cases of excavations supported by floating walls, the effect of embedded wall length, depth of the stiff layer, bending stiffness of wall and excavation geometry and over-consolidation ratio of soils were found to have a influence on the maximum wall deflection. In general, the predictions fell within 30% of the finite element computed results. A new chart Ψ versus normalized system stiffness was used to demonstrate that MSD could correctly capture the trend of wall displacements increasing with the ratio ofexcavation depth to depth of stiff layer, which could be controlled by increasing wall stiffness for very stiff wall system only. The incorporation of a simple parabolic curvequantifying small strain stiffness of soil was proven to be essential to good ground movement predictions. A new dimensionless group has been defined using the MSD concepts to analyze 110 cases of excavation. The new database can now be used to investigate the relationship between structural response ratio S and soil-structure stiffness ratio R where this is shown on log-log axes to capture the enormous range of wall stiffness between sheet-piles and thick diaphragm walls. Wall stiffness was found to have a negligible influence on the magnitude of the wall bulging displacements for deep excavation supported by fixed-based wall with stiffness ranging from sheet pile walls to ordinary reinforced concrete diaphragm walls, whereas excavations supported by floating walls were found to be influenced by wall stiffness due to the difference in deformation mechanisms.