No description available.

No description available.

No description available.

Matting, which aims to separate the foreground object from the background of an image, is an important problem in computer vision. Most existing methods rely on auxiliary information such as trimaps or scibbles to alleviate the difficulty arising from the underdetermined nature of the matting problem. However, such methods tend to be sensitive to the quality of auxiliary information, and are unsuitable for real-time deployment. In this paper, we propose a novel Attention-based Multi-scale Matting Network (AMMNet), which can estimate the alpha matte from a given RGB image without resorting to any auxiliary information. The proposed AMMNet consists of three (sub-)networks: 1) a multi-scale neural network designed to provide the semantic information of the foreground object, 2) a Unet-like network for attention mask generation, and 3) a Convolutional Neural Network (CNN) customized to integrate high- and low-level features extracted by the first two (sub-)networks. The AMMNet is generic in nature and can be trained end-to-end in a straightforward manner. The experimental results indicate that the performance of AMMNet is competitive against the state-of-the-art matting methods, which either require additional side information or are tailored to images with a specific type of content (e.g., portrait). / Thesis / Master of Applied Science (MASc)

One of the major problems in computer vision involves dealing with uncertain information. Occlusion, dissimilar views, insufficient illumination, insufficient resolution, and degradation give rise to imprecise data. At the same time, incomplete or local knowledge of the scene gives rise to imprecise interpretation rules. Uncertainty arises at different processing levels of computer vision either because of the imprecise data or because of the imprecise interpretation rules. It is natural to build computer vision systems that incorporate uncertainty reasoning. The Dempster-Shafer (D-S) theory of evidence is appealing for coping with uncertainty hierarchically. However, very little work has been done to apply D-S theory to practical vision systems because some important problems are yet to be resolved. / Ph. D.

The purpose of this dissertation is to present our agent-based human tracking framework, and to evaluate the results of our work in light of the previous research in the same field. Our agent-based approach departs from a process-centric model where the agents are bound to specific processes, and introduces a novel model by which agents are bound to the objects or sub-objects being recognized or tracked. The hierarchical agent-based model allows the system to handle a variety of cases, such as single people or multiple people in front of single or stereo cameras. We employ the job-market model for agents' communication. In this dissertation, we will present several experiments in detail, which demonstrate the effectiveness of the agent-based tracking system. Per our research, the agents are designed to be autonomous, self-aware entities that are capable of communicating with other agents to perform tracking within agent coalitions. Each agent with high-level abstracted knowledge seeks evidence for its existence from the low-level features (e.g. motion vector fields, color blobs) and its peers (other agents representing body-parts with which it is compatible). The power of the agent-based approach is its flexibility by which the domain information may be encoded within each agent to produce an overall tracking solution. / Ph. D.

Painting still plays a fundamental role in communication nowadays. For example, the paint on the road, called road surface marking, guides the traffic in order and maintains the high efficiency of the entire modern traffic system. With the development of the Autonomous Ground Vehicle (AGV), the idea of a line Painting Robot emerged. In this thesis, a Painting Robot was designed as a standalone system based on the AGV platform. In this study, the mechanical and electronic design of a Painting Robot was discussed. The overall design was to fulfill the requirements of the line painting. Computer vision techniques were applied to this thesis since the camera was selected as the major sensor of the robot. Advanced control theory was introduced to this thesis as well. Three different controllers were developed. The Proportional-Integral (PI) controller with an anti-windup feature was designed to overcome the drawbacks of the traditional PI controller. Model Reference Adaptive Control (MRAC) was introduced into this thesis to deal with the uncertainties of the system. At last, the hybrid PI-MRAC controller was implemented to maintain the advantages of both PI and MRAC approaches. Experiments were conducted to evaluate the performance of the entire system, which indicated the successful design of the Painting Robot. / Master of Science / Painting still plays a fundamental role in communication nowadays. With the development of the Autonomous Ground Vehicle (AGV), the idea of a line Painting Robot emerged. In this thesis, a Painting Robot was designed as a standalone system based on the AGV platform. In this study, a Painting Robot with a two-camera system was designed. Computer vision techniques and advanced control theory were introduced into this thesis. Three different controllers were developed, including Proportional-Integral (PI) with an anti-windup feature, Model Reference Adaptive Control (MRAC) and the hybrid PI-MRAC. Experiments were conducted to evaluate the performance of the entire system, which indicated the successful design of the Painting Robot.

The ability to quickly, accurately, and autonomously identify and track objects in digital images in real-time has been an area of investigation for quite some time. Research in this area falls under the broader category of computer vision. Only in recent decades, with advances in computing power and commercial optical hardware, has this capability become a possibility. There are many different methods of identifying and tracking objects of interest, and best practices are still being developed, varying based on application. This thesis examines background subtraction methods as they apply to the tracking of small unmanned aerial systems (sUAS). A system combining commercial off-the-shelf (COTS) cameras and a pan-tilt unit (PTU), along with custom developed code, is developed for the purpose of continuously pointing at and tracking the motion of a sUAS in flight. Mixtures of Gaussians Background Modeling (MOGBM) is used to track the motion of the sUAS in frame and determine when to command the PTU. When the camera is moving, background subtraction methods are unusable, so additional methods are explored for filling this performance gap. The stereo vision capabilities of the system, enabled by the use of two cameras simultaneously, allow for estimation of the three-dimensional position and trajectory of the sUAS. This system can be used as a supplement or replacement to traditional tracking methods such as GPS and RADAR as part of a larger unmanned aerial systems traffic control (UTC) infrastructure. / Master of Science / The ability to quickly, accurately, and automatically identify and track targets in digital images has been of interest for some time now. Research in this area falls under the broader category of computer vision. Only in recent decades, with advances in computing power and commercial optical hardware, has this ability become a possibility. There are many different methods of identifying and tracking targets of interest, and best practices are still being developed, varying based on application. This thesis examines background subtraction methods as they apply to the tracking of small unmanned aerial systems (sUAS), commonly referred to as drones. A system combining cameras and a moving platform, along with custom developed code, is developed for the purpose of continuously pointing at and tracking the motion of an sUAS in flight. The system is able to map out the three-dimensional position of a flying sUAS over time.

Neural networks in their modern deep learning incarnation have achieved state of the art performance on a wide variety of tasks and domains. A core intuition behind these methods is that they learn layers of features which interpolate between two domains in a series of related parts. The first part of this thesis introduces the building blocks of neural networks for computer vision. It starts with linear models then proceeds to deep multilayer perceptrons and convolutional neural networks, presenting the core details of each. However, the introduction also focuses on intuition by visualizing concrete examples of the parts of a modern network. The second part of this thesis investigates regularization of neural networks. Methods like dropout and others have been proposed to favor certain (empirically better) solutions over others. However, big deep neural networks still overfit very easily. This section proposes a new regularizer called DeCov, which leads to significantly reduced overfitting (difference between train and val performance) and greater generalization, sometimes better than dropout and other times not. The regularizer is based on the cross-covariance of hidden representations and takes advantage of the intuition that different features should try to represent different things, an intuition others have explored with similar losses. Experiments across a range of datasets and network architectures demonstrate reduced overfitting due to DeCov while almost always maintaining or increasing generalization performance and often improving performance over dropout. / Master of Science

Small unmanned aircraft systems (UAS) must be able to detect and avoid conflicting traffic, an especially challenging task when the threat is another small UAS. Collision avoidance requires trajectory prediction and the performance of a collision avoidance system can be improved by extending the prediction horizon. In this thesis, an algorithm for predicting the trajectory of a small, fixed-wing UAS using an estimate of its orientation and for maneuvering around the threat, if necessary, is developed. A computer vision algorithm locates specific feature points of a threat aircraft in an image and the POSIT algorithm uses these feature points to estimate the pose (position and attitude) of the threat. A sequence of pose estimates is then used to predict the trajectory of the threat aircraft and to avoid colliding with it. To assess the algorithm's performance, the predictions are compared with predictions based solely on position estimates for a variety of encounter scenarios. Simulation and experimental results indicate that trajectory prediction using orientation estimates provides quicker response to a change in the threat aircraft trajectory and results in better prediction and avoidance performance. / Master of Science