With the recent advancement of deep learning, the performance of object detection techniques has greatly increased in both speed and accuracy. This has made it possible to run highly accurate object detection with real time speed on modern desktop computer systems. Recently, there has been a growing interest in developing smaller and faster deep neural network architectures suited for embedded devices. This thesis explores the suitability of running object detection on the Raspberry Pi 3, a popular embedded computer board. Two controlled experiments are conducted where two state of the art object detection models SSD and YOLO are tested in how they perform in accuracy and speed. The results show that the SSD model slightly outperforms YOLO in both speed and accuracy, but with the low processing power that the current generation of Raspberry Pi has to offer, none of the two performs well enough to be viable in applications where high speed is necessary.

Advances in medical imaging technology have led to the acquisition of large number of images in different modalities. On some of these images the boundaries of key organs need to be accurately identified for treatment planning and diagnosis. This is typically performed manually by a physician who uses prior knowledge of organ shapes and locations to demarcate the boundaries of organs. Such manual segmentation is subjective, time consuming and prone to inconsistency. Automating this task has been found to be very challenging due to poor tissue contrast and ill-defined organ/tissue boundaries. This dissertation presents a genetic algorithm for combining representations of learned information such as known shapes, regional properties and relative location of objects into a single framework in order to perform automated segmentation. The algorithm has been tested on two different datasets: for segmenting hands on thermographic images and for prostate segmentation on pelvic computed tomography (CT) and magnetic resonance (MR) images. In this dissertation we report the results of segmentation in two dimensions (2D) for thermographic images; and two as well as three dimensions (3D) for pelvic images. We show that combining multiple features for segmentation improves segmentation accuracy as compared with segmentation using single features such as texture or shape alone.

I conduct a study of learning in HMAX-like models, which are hierarchical models of visual processing in biological vision systems. Such models compute a new representation for an image based on the similarity of image sub-parts to a number of specific patterns, called prototypes. Despite being a central piece of the overall model, the issue of choosing the best prototypes for a given task is still an open problem. I study this problem, and consider the best way to increase task performance while decreasing the computational costs of the model. This work broadens our understanding of HMAX and related hierarchical models as tools for theoretical neuroscience, while simultaneously increasing the utility of such models as applied computer vision systems.

In this thesis, I explore a reinforcement learning technique for improving bounding box localizations of objects in images. The model takes as input a bounding box already known to overlap an object and aims to improve the fit of the box through a series of transformations that shift the location of the box by translation, or change its size or aspect ratio. Over the course of these actions, the model adapts to new information extracted from the image. This active localization approach contrasts with existing bounding-box regression methods, which extract information from the image only once. I implement, train, and test this reinforcement learning model using data taken from the Portland State Dog-Walking image set. The model balances exploration with exploitation in training using an ε-greedy policy. I find that the performance of the model is sensitive to the ε-greedy configuration used during training, performing best when the epsilon parameter is set to very low values over the course of training. With = 0.01, I find the algorithm can improve bounding boxes in about 78% of test cases for the "dog" object category, and 76% for the "human" category.

Recent work has shown that flow estimation from a pair of images can be formulated as a supervised learning task to be resolved with convolutional neural networks (CNN). However, the basic straightforward CNN methods estimate optical flow with motion and occlusion boundary blur. To tackle this problem, we propose a tiny diagnostic dataset called FlowClevr to quickly evaluate various modules that can use to enhance standard CNN architectures. Based on the experiments of the FlowClevr dataset, we find that a deformable module can improve model prediction accuracy by around 30% to 100% in most tasks and more significantly reduce boundary blur. Based on these results, we are able to design modifications to various existing network architectures improving their performance. Compared with the original model, the model with the deformable module clearly reduces boundary blur and achieves a large improvement on the MPI sintel dataset, an omni-directional stereo (ODS) and a novel omni-directional optical flow dataset.

Orthopaedic residency training is in the midst of a paradigm shift. Recent mandates from the Residency Review Committee (RRC) for Orthopaedic Surgery and the American Board of Orthopaedic Surgery (ABOS) are requiring that programs must provide structured motor skills training to first year residents. Although other surgical fields such as laparoscopic surgery have been using simulation tools to train incoming residents for over a decade, the orthopaedic field has lagged behind in developing these training tools. Given the need for orthopaedic training devices and the lack of currently available solutions to residency programs, this work has focused on developing a surgical simulator for the task of hip guide wire navigation. Hip wire navigation was targeted for this work because it is a core competency skill for surgical residents and few options currently exist for training residents on this task. Much of this work focuses on the development of the wire navigation simulator. The simulator has six main components; a single camera interfaced with a Raspberry Pi (a credit-card sized computer), a series of three mirrors, a surrogate femur, a guide wire driver, a laser etched guide wire, and a laptop. These components interact to create virtual radiograph images that the resident can use to place the guide wire inside the bone. The goal in developing this simulator is to provide a platform which enables residents to acquire the skill of hip wire navigation in a safe environment and eventually transfer that skill into the operating room. Assessment of the simulator has shown that the guide wire can be located in bone within 1.5mm of its true position and less than a degree of its true trajectory. This level of accuracy is sufficient for providing residents with a training tool to practice their technique on. In training with resident surgeons, initial trends show that practicing with the simulator can result in an improvement in one’s technique. Residents who have trained with the simulator show a decrease in both the amount of radiographic images required to complete the procedure and the amount of time required to perform the procedure in a pseudo operating room environment. While more work is needed to be done to show the significance of this trend, this work has achieved its goal of providing residents with a safe platform for practicing the task of hip guide wire navigation.

In the study of computer vision, background modeling is a fundamental and critical task in many conventional applications. This thesis presents an introduction to background modeling and various computer vision techniques for estimating the background model to achieve the goal of removing dynamic objects in a video sequence. The process of estimating the background model with temporal changes in the absence of foreground moving objects is called adaptive background modeling. In this thesis, three adaptive background modeling approaches were presented for the purpose of developing \teacher removal" algorithms. First, an adaptive background modeling algorithm based on linear adaptive prediction is presented. Second, an adaptive background modeling algorithm based on statistical dispersion is presented. Third, a novel adaptive background modeling algorithm based on low rank and sparsity constraints is presented. The design and implementation of these algorithms are discussed in detail, and the experimental results produced by each algorithm are presented. Lastly, the results of this research are generalized and potential future research is discussed.

No description available.

No description available.

With the advancement of computer technology, demand for more accurate and intelligent monitoring systems has also risen. The use of computer vision and video analysis range from industrial inspection to surveillance. Object detection and segmentation are the first and fundamental task in the analysis of dynamic scenes. Traditionally, this detection and segmentation are typically done through temporal differencing or statistical modelling methods. One of the most widely used background modeling and segmentation algorithms is the Mixture of Gaussians method developed by Stauffer and Grimson (1999). During the past decade many such algorithms have been developed ranging from parametric to non-parametric algorithms. Many of them utilise pixel intensities to model the background, but some use texture properties such as Local Binary Patterns. These algorithms function quite well under normal environmental conditions and each has its own set of advantages and short comings. However, there are two drawbacks in common. The first is that of the stationary object problem; when moving objects become stationary, they get merged into the background. The second problem is that of light changes; when rapid illumination changes occur in the environment, these background modelling algorithms produce large areas of false positives. / These algorithms are capable of adapting to the change, however, the quality of the segmentation is very poor during the adaptation phase. In this thesis, a framework to suppress these false positives is introduced. Image properties such as edges and textures are utilised to reduce the amount of false positives during adaptation phase. The framework is built on the idea of sequential pattern recognition. In any background modelling algorithm, the importance of multiple image features as well as different spatial scales cannot be overlooked. Failure to focus attention on these two factors will result in difficulty to detect and reduce false alarms caused by rapid light change and other conditions. The use of edge features in false alarm suppression is also explored. Edges are somewhat more resistant to environmental changes in video scenes. The assumption here is that regardless of environmental changes, such as that of illumination change, the edges of the objects should remain the same. The edge based approach is tested on several videos containing rapid light changes and shows promising results. Texture is then used to analyse video images and remove false alarm regions. Texture gradient approach and Laws Texture Energy Measures are used to find and remove false positives. It is found that Laws Texture Energy Measure performs better than the gradient approach. The results of using edges, texture and different combination of the two in false positive suppression are also presented in this work. This false positive suppression framework is applied to a smart house senario that uses cameras to model ”virtual sensors” to detect interactions of occupants with devices. Results show the accuracy of virtual sensors compared with the ground truth is improved.