1151 |
BOR2G : Building Optimal Regularised Reconstructions with GPUs (in cubes)Tanner, Michael January 2017 (has links)
Robots require high-quality maps - internal representations of their operating workspace - to localise, path plan, and perceive their environment. Until recently, these maps were restricted to sparse, 2D representations due to computational, memory, and sensor limitations. With the widespread adoption of high-quality sensors and graphics processors for parallel processing, these restrictions no longer apply: dense 3D maps are feasible to compute in real time (i.e., at the input sensor's frame rate). This thesis presents the theory and system to create large-scale dense 3D maps (i.e., reconstruct continuous surface models) using only sensors found on modern autonomous automobiles: 2D laser, 3D laser, and cameras. In contrast to active RGB-D cameras, passive cameras produce noisy surface observations and must be regularised in both 2D and 3D to create accurate reconstructions. Unfortunately, straight-forward application of 3D regularisation causes undesired surface interpolation and extrapolation in regions unexplored by the robot. We propose a method to overcome this challenge by informing the regulariser of the specific subsets of 3D surfaces upon which to operate. When combined with a compressed voxel grid data structure, we demonstrate our system fusing data from both laser and camera sensors to reconstruct 7.3 km of urban environments. We evaluate the quantitative performance of our proposed method through the use of synthetic and real-world datasets - including datasets from Stanford's Burghers of Calais, University of Oxford's RobotCar, University of Oxford's Dense Reconstruction, and Karlsruhe Institute of Technology's KITTI - compared to ground-truth laser data. With only stereo camera inputs, our regulariser reduces the 3D reconstruction metric error between 27% to 36% with a final median accuracy ranging between 4 cm to 8 cm. Furthermore, by augmenting our system with object detection, we remove ephemeral objects (e.g., automobiles, bicycles, and pedestrians) from the input sensor data and target our regulariser to interpolate the occluded urban surfaces. Augmented with Kernel Conditional Density Estimation, our regulariser creates reconstructions with median errors between 5.64 cm and 9.24 cm. Finally, we present a machine-learning pipeline that learns, in an automatic fashion, to recognise the errors in dense reconstructions. Our system trains on image and laser data from a 3.8 km urban sequence. Using a separate 2.2 km urban sequence, our pipeline consistently identifies error-prone regions in the image-based dense reconstruction.
|
1152 |
Automated alarm and root-cause analysis based on real time high-dimensional process data : Part of a joint research project between UmU, Volvo AB & Volvo CarsHarbs, Justin, Svensson, Jack January 2018 (has links)
Today, a large amount of raw data are available within manufacturing industries. Unfortunately, most of it is not further analyzed in search of valuable information regarding the optimization of processes. In the painting process at the Volvo plant in Umeå, adjusted settings on the process equipments (e.g. robots, machines etc.) are mostly based on the experience of the personnel rather than actual facts (i.e. analyzed data). Consequently, time- and cost waste caused by defects is obtained when painting the commercial heavy-duty truck bodies (cabs). Hence, the aim of this masters thesis is to model the quality as a function of available background- and process data. This should be presented in an automated alarm and root-cause system. A variety of supervised learning algorithms were trained in order to estimate the probability of having at least one defect per cab. Even with a small amount of data, results have shown that such algorithms can provide valuable information. Later in this thesis work, one of the algorithms was chosen and used as the underlying model in the prototype of an automated alarm system. When this probability was considered as too high, an intuitive root-cause analysis was presented. Ultimately, this research has demonstrated the importance and possibility of analyzing data with statistical tools in the search of limiting costs- and time waste.
|
1153 |
Diagnosis of autonomous vehicles using machine learningHossain, Adnan January 2018 (has links)
With autonomous trucks on the road where the driver is absent requires new diagnostic methods. The driver possess several abilities which a machine does not. In this thesis, the use of machine learning as a method was investigated. A more concrete problem description was formed where the main objective was detecting anomalies in wheel configurations. More specifically, the machine learning model was used to detect incorrect wheel settings. Three different algorithms was used, SVM, LDA and logistic regression. Overall, the classifier predicts with high accuracy supporting that machine learning can be used for diagnosing autonomous vehicles.
|
1154 |
On Bayesian optimization and its application to hyperparameter tuningMatosevic, Antonio January 2018 (has links)
This thesis introduces the concept of Bayesian optimization, primarly used in optimizing costly black-box functions. Besides theoretical treatment of the topic, the focus of the thesis is on two numerical experiments. Firstly, different types of acquisition functions, which are the key components responsible for the performance, are tested and compared. Special emphasis is on the analysis of a so-called exploration-exploitation trade-off. Secondly, one of the most recent applications of Bayesian optimization concerns hyperparameter tuning in machine learning algorithms, where the objective function is expensive to evaluate and not given analytically. However, some results indicate that much simpler methods can give similar results. Our contribution is therefore a statistical comparison of simple random search and Bayesian optimization in the context of finding the optimal set of hyperparameters in support vector regression. It has been found that there is no significant difference in performance of these two methods.
|
1155 |
Reducing animator keyframesHolden, Daniel January 2017 (has links)
The aim of this doctoral thesis is to present a body of work aimed at reducing the time spent by animators manually constructing keyframed animation. To this end we present a number of state of the art machine learning techniques applied to the domain of character animation. Data-driven tools for the synthesis and production of character animation have a good track record of success. In particular, they have been adopted thoroughly in the games industry as they allow designers as well as animators to simply specify the high-level descriptions of the animations to be created, and the rest is produced automatically. Even so, these techniques have not been thoroughly adopted in the film industry in the production of keyframe based animation [Planet, 2012]. Due to this, the cost of producing high quality keyframed animation remains very high, and the time of professional animators is increasingly precious. We present our work in four main chapters. We first tackle the key problem in the adoption of data-driven tools for key framed animation - a problem called the inversion of the rig function. Secondly, we show the construction of a new tool for data-driven character animation called the motion manifold - a representation of motion constructed using deep learning that has a number of properties useful for animation research. Thirdly, we show how the motion manifold can be extended as a general tool for performing data-driven animation synthesis and editing. Finally, we show how these techniques developed for keyframed animation can also be adapted to advance the state of the art in the games industry.
|
1156 |
User-centric traffic engineering in software defined networksBakhshi, Taimur January 2017 (has links)
Software defined networking (SDN) is a relatively new paradigm that decouples individual network elements from the control logic, offering real-time network programmability, translating high level policy abstractions into low level device configurations. The framework comprises of the data (forwarding) plane incorporating network devices, while the control logic and network services reside in the control and application planes respectively. Operators can optimize the network fabric to yield performance gains for individual applications and services utilizing flow metering and application-awareness, the default traffic management method in SDN. Existing approaches to traffic optimization, however, do not explicitly consider user application trends. Recent SDN traffic engineering designs either offer improvements for typical time-critical applications or focus on devising monitoring solutions aimed at measuring performance metrics of the respective services. The performance caveats of isolated service differentiation on the end users may be substantial considering the growth in Internet and network applications on offer and the resulting diversity in user activities. Application-level flow metering schemes therefore, fall short of fully exploiting the real-time network provisioning capability offered by SDN instead relying on rather static traffic control primitives frequent in legacy networking. For individual users, SDN may lead to substantial improvements if the framework allows operators to allocate resources while accounting for a user-centric mix of applications. This thesis explores the user traffic application trends in different network environments and proposes a novel user traffic profiling framework to aid the SDN control plane (controller) in accurately configuring network elements for a broad spectrum of users without impeding specific application requirements. This thesis starts with a critical review of existing traffic engineering solutions in SDN and highlights recent and ongoing work in network optimization studies. Predominant existing segregated application policy based controls in SDN do not consider the cost of isolated application gains on parallel SDN services and resulting consequence for users having varying application usage. Therefore, attention is given to investigating techniques which may capture the user behaviour for possible integration in SDN traffic controls. To this end, profiling of user application traffic trends is identified as a technique which may offer insight into the inherent diversity in user activities and offer possible incorporation in SDN based traffic engineering. A series of subsequent user traffic profiling studies are carried out in this regard employing network flow statistics collected from residential and enterprise network environments. Utilizing machine learning techniques including the prominent unsupervised k-means cluster analysis, user generated traffic flows are cluster analysed and the derived profiles in each networking environment are benchmarked for stability before integration in SDN control solutions. In parallel, a novel flow-based traffic classifier is designed to yield high accuracy in identifying user application flows and the traffic profiling mechanism is automated. The core functions of the novel user-centric traffic engineering solution are validated by the implementation of traffic profiling based SDN network control applications in residential, data center and campus based SDN environments. A series of simulations highlighting varying traffic conditions and profile based policy controls are designed and evaluated in each network setting using the traffic profiles derived from realistic environments to demonstrate the effectiveness of the traffic management solution. The overall network performance metrics per profile show substantive gains, proportional to operator defined user profile prioritization policies despite high traffic load conditions. The proposed user-centric SDN traffic engineering framework therefore, dynamically provisions data plane resources among different user traffic classes (profiles), capturing user behaviour to define and implement network policy controls, going beyond isolated application management.
|
1157 |
Multi-Task Learning via Structured Regularization: Formulations, Algorithms, and ApplicationsJanuary 2011 (has links)
abstract: Multi-task learning (MTL) aims to improve the generalization performance (of the resulting classifiers) by learning multiple related tasks simultaneously. Specifically, MTL exploits the intrinsic task relatedness, based on which the informative domain knowledge from each task can be shared across multiple tasks and thus facilitate the individual task learning. It is particularly desirable to share the domain knowledge (among the tasks) when there are a number of related tasks but only limited training data is available for each task. Modeling the relationship of multiple tasks is critical to the generalization performance of the MTL algorithms. In this dissertation, I propose a series of MTL approaches which assume that multiple tasks are intrinsically related via a shared low-dimensional feature space. The proposed MTL approaches are developed to deal with different scenarios and settings; they are respectively formulated as mathematical optimization problems of minimizing the empirical loss regularized by different structures. For all proposed MTL formulations, I develop the associated optimization algorithms to find their globally optimal solution efficiently. I also conduct theoretical analysis for certain MTL approaches by deriving the globally optimal solution recovery condition and the performance bound. To demonstrate the practical performance, I apply the proposed MTL approaches on different real-world applications: (1) Automated annotation of the Drosophila gene expression pattern images; (2) Categorization of the Yahoo web pages. Our experimental results demonstrate the efficiency and effectiveness of the proposed algorithms. / Dissertation/Thesis / Ph.D. Computer Science 2011
|
1158 |
Single atom imaging with time-resolved electron microscopyFurnival, Thomas January 2017 (has links)
Developments in scanning transmission electron microscopy (STEM) have opened up new possibilities for time-resolved imaging at the atomic scale. However, rapid imaging of single atom dynamics brings with it a new set of challenges, particularly regarding noise and the interaction between the electron beam and the specimen. This thesis develops a set of analytical tools for capturing atomic motion and analyzing the dynamic behaviour of materials at the atomic scale. Machine learning is increasingly playing an important role in the analysis of electron microscopy data. In this light, new unsupervised learning tools are developed here for noise removal under low-dose imaging conditions and for identifying the motion of surface atoms. The scope for real-time processing and analysis is also explored, which is of rising importance as electron microscopy datasets grow in size and complexity. These advances in image processing and analysis are combined with computational modelling to uncover new chemical and physical insights into the motion of atoms adsorbed onto surfaces. Of particular interest are systems for heterogeneous catalysis, where the catalytic activity can depend intimately on the atomic environment. The study of Cu atoms on a graphene oxide support reveals that the atoms undergo anomalous diffusion as a result of spatial and energetic disorder present in the substrate. The investigation is extended to examine the structure and stability of small Cu clusters on graphene oxide, with atomistic modelling used to understand the significant role played by the substrate. Finally, the analytical methods are used to study the surface reconstruction of silicon alongside the electron beam-induced motion of adatoms on the surface. Taken together, these studies demonstrate the materials insights that can be obtained with time-resolved STEM imaging, and highlight the importance of combining state-ofthe- art imaging with computational analysis and atomistic modelling to quantitatively characterize the behaviour of materials with atomic resolution.
|
1159 |
Relationship descriptors for interactive motion adaptationAl-Ashqar, Rami January 2017 (has links)
In this thesis we present an interactive motion adaptation scheme for close interactions between skeletal characters and mesh structures, such as navigating restricted environments and manipulating tools. We propose a new spatial-relationship based representation to encode character-object interactions describing the kinematics of the body parts by the weighted sum of vectors relative to descriptor points selectively sampled over the scene. In contrast to previous discrete representations that either only handle static spatial relationships, or require offline, costly optimization processes, our continuous framework smoothly adapts the motion of a character to deformations in the objects and character morphologies in real-time whilst preserving the original context and style of the scene. We demonstrate the strength of working in our relationship-descriptor space in tackling the issue of motion editing under large environment deformations by integrating procedural animation techniques such as repositioning contacts in an interaction whilst preserving the context and style of the original animation. Furthermore we propose a method that can be used to adapt animations from template objects to novel ones by solving for mappings between the two in our relationship-descriptor space effectively transferring an entire motion from one object to a new one of different geometry whilst ensuring continuity across all frames of the animation, as opposed to mapping static poses only as is traditionally achieved. The experimental results show that our method can be used for a wide range of applications, including motion retargeting for dynamically changing scenes, multi-character interactions, and interactive character control and deformation transfer for scenes that involve close interactions. We further demonstrate a key use case in retargeting locomotion to uneven terrains and curving paths convincingly for bipeds and quadrupeds. Our framework is useful for artists who need to design animated scenes interactively, and modern computer games that allow users to design their own virtual characters, objects and environments, such that they can recycle existing motion data for a large variety of different configurations without the need to manually reconfigure motion from scratch or store expensive combinations of animation in memory. Most importantly it’s achieved in real-time.
|
1160 |
Towards the Design of Neural Network Framework for Object Recognition and Target Region Refining for Smart Transportation SystemsZhao, Yiheng 13 August 2018 (has links)
Object recognition systems have significant influences on modern life. Face, iris and finger point recognition applications are commonly applied for the security purposes; ASR (Automatic Speech Recognition) is commonly implemented on speech subtitle generation for various videos and audios, such as YouTube; HWR (Handwriting Recognition) systems are essential on the post office for cheque and postcode detection; ADAS (Advanced Driver Assistance System) are well applied to improve drivers’, passages’ and pedestrians’ safety. Object recognition techniques are crucial and valuable for academia, commerce and industry.
Accuracy and efficiency are two important standards to evaluate the performance of recognition techniques. Accuracy includes how many objects can be indicated in real scene and how many of them can be correctly classified. Efficiency means speed for system training and sample testing. Traditional object detecting methods, such as HOG (Histogram of orientated Gradient) feature detector combining with SVM (Support Vector Machine) classifier, cannot compete with frameworks of neural networks in both efficiency and accuracy. Since neural network has better performance and potential for improvement, it is worth to gain insight into this field to design more advanced recognition systems.
In this thesis, we list and analyze sophisticated techniques and frameworks for object recognition. To understand the mathematical theory for network design, state-of-the-art networks in ILSVRC (ImageNET Large Scale Visual Recognition Challenge) are studied. Based on analysis and the concept of edge detectors, a simple CNN (Convolutional Neural Network) structure is designed as a trail to explore the possibility to utilize the network of high width and low depth for region proposal selection, object recognition and target region refining. We adopt Le-Net as the template, taking advantage of multi-kernels of GoogLe-Net.
We made experiments to test the performance of this simple structure of the vehicle and face through ImageNet dataset. The accuracy for the single object detection is 81% in average and for plural object detection is 73.5%. We refined networks through many aspects to reach the final accuracy 95% for single object detection and 89% for plural object detection.
|
Page generated in 0.112 seconds