教学课程资源库 更多>>

关键词： Circuits   Algorithms   Graph coloring   Computer science   Electrical engineering   Engineering   Mathematics  

关键词： Robotics   Velocity   Data collection   Datasets   Experiments   Calibration   Sensors   Electromagnetics   Computer science  

摘要： Autonomous mobile robots are being deployed in ever more varied roles and environments. For instance, a robotic explorer was deployed just this year to Jezero Crater, the most challenging area of Mars yet explored. This might lead one to believe autonomous mobile robots are able to operate in almost any environment. But in fact there exists a broad range of environments that are inaccessible to autonomous robots. Environments with smoke, fog, darkness, and other visually challenging conditions, collectively referred to as visually-degraded environments (VDEs), are among the biggest hurdles to deploying autonomous mobile robots in the field. But a robot is likely to encounter such conditions almost anywhere outside of a well-lit indoor environment. So if autonomous mobile robots are to be truly useful in a broad range of environments, the problem of perception in VDEs must be addressed. In theory, one could avoid visual challenges simply by using a different type of sensor. Millimeter wave radar, for instance, is unaffected by most kinds of airborne particulates and it does not depend on ambient light. But using radar for robotic perception comes with its own challenges. Popular visual and lidar-based methods for perception and state estimation do not adapt well to radar. Additionally, radar measurements are subject to forms of noise unknown in other sensors. This dissertation will cover a number of novel developments in perception and state estimation using millimeter wave radar that address these issues including i) a radar-inertial method for state estimation in smoke and fog, ii) a radar-only occupancy mapping method, iii) a unique dataset for radar-based state estimation and perception, iv) an end-to-end learned method for 3D radar image alignment, and v) a new radar-only detector for moving obstacles in road environments.

关键词： Computer science   Mathematics   Cryptography   Construction   Random variables   Bandwidths   Concrete   Dissertations & theses   Polynomials   Codes   Algorithms   Heuristic   Energy consumption   Reading  

摘要： Nowadays, attackers have large amount of resources (e.g., computational and space capabilities) that are used for malicious purposes, e.g., stealing sensitive information (such as passwords). To face these capacious adversaries, “resource-demanding” functions have been developed: Functions that deliberately consume (on evaluation) large quantities of resources (e.g., time, space, energy, etc.). In this light, a recent work of Boneh et al. (CRYPTO18) introduces the notion of verifiable delay function, a time-demanding function that requires a minimum number of steps (time complexity) in order to be evaluated. The main property, that puts verifiable delay functions apart from the state of the art resource-demanding functions, is the possibility of checking the correctness of the computation efficiently, i.e., the time complexity of the verification process is significantly smaller than the time complexity required in order to evaluate the function. In this dissertation, we investigate further the area of verifiable resource-demanding functions by proposing the notion of verifiable capacity-bound function (VCBF). The main VCBF property is that it imposes a lower bound on the number of bits read from memory during evaluation (referred to as minimum capacity). No adversary, even with unbounded resources, should produce an output without spending this minimum memory capacity. Moreover, a VCBF allows for an efficient public verification process: Given a proof-of-correctness, checking the validity of the output takes significantly fewer memory resources, sublinear in the target minimum capacity. Finally, it achieves soundness, i.e., no computationally bounded adversary can produce a proof that passes verification for a false output. With these properties, we believe a VCBF can be viewed as a “space” analog of a verifiable delay function. We then propose the first VCBF construction relying on evaluating a degree-d polynomial f from Fp[x] at a random point. To prove the space

关键词： Artificial intelligence   Computer science   Computer engineering   Sparsity   Motivation   Accuracy   Cellular telephones   Mutation   Neural networks   Bias  

摘要： Convolution filters in CNNs extract patterns from input by aggregating information across height, width and channel dimensions. Information aggregation across height and width dimensions performed using depthwise convolution, helps identify neighborhood patterns and hence is very intuitive. However the method in which channel dimension information is aggregated by channel summation seems mathematically simplistic and out of convenience. In this project we attempt to improve the channel dimension aggregation operations. The first approach introduces weighted summation channel aggregation in convolutions. The second approach introduces permuted convolutions which attempt to perform psuedo-width scaling by generating new constrained filters from existing filters. Implementing permuted convolutions comes with many challenges such as permutation explosion, stochasticity, higher memory and computation requirements. To resolve these issues, we come up with multiple variants of permuted convolutions and present their advantages and disadvantages. Lastly, we provide empirical results showcasing the performance of weighted channel summation networks and permuted convolution networks, present our findings and recommendations for future work.

关键词： Artificial intelligence   Computer science  

摘要： In the development of science, humans learned from nature by taxonomical approaches. We named, defined, and classified the natural phenomena to understand underlying principles. Especially, the classification is a key step to extract common features from resemblant objects or incidents. As the automata arise in various fields, automation of classification tasks is extensively studied. Some classification tasks, in which underlying principles are expressed in the computable mathematical form, are easily executed by machines. However, the other classification tasks, in which underlying principles are not understood completely, are still under investigation. Interestingly, the collection of artificial neurons inspired by the biological neurons successfully classifies images without having any knowledge about the underlying principles. The deep neural networks also show outstanding performances on other systems such as reinforcement learning, driving automation, etc. even though we do not understand the systems themselves due to the complication of the problems. In this dissertation, we investigate the underlying principles behind neural networks via input space. We can understand how and why neural networks work via the input space since a human can intuitively explain the input. Especially, the piece-wise linear neural networks have a unique mathematical property that divides the input space into numerous linear regions, and each linear region shares the same gradient with respect to the input. Based on this property, we can explain why the neural networks generalize better than any other methods even though they contain numerous parameters and why the training neural network is successful using the cross-entropy loss. We also suggest a novel loss function that can be used for training the neural network as well. Furthermore, we introduce a novel importance score, named as influential score, to interpret the output of neural networks by tracing the significant units i

关键词： Artificial intelligence   Computer science   Disability studies   Mechanics   Physics  

摘要： Gait analysis is typically conducted using an optoelectronic system which is known as the standard method for motion analysis. Despite advance development of instruments related to the optoelectronic approach, there are still a few limitations of the traditional gait analysis which limit the accessibility for individuals who would benefit from the investigation. A newly developed three-dimension motion capture system, known as Inertial Measurement Units (IMU) was introduced as an option for gait analysis. The IMU system is a transportable camera-free motion capture system. This also motivated the principle of out-of-the lab gait analysis. To broaden the use of the new system, this PhD project was conducted to examine whether the system should be used confidently for clinical gait analysis. The main purpose of this PhD project was to examine the feasibility of incorporating a machine learning method to estimate the kinetics of gait using the kinematics data obtained from an IMU system. Firstly, as pilot studies, an artificial neural network (ANN) was trained using gait data derived from the potential input signals which were signals of marker coordinates and joint angles obtained from an IMU system (Xsens) to predict joint moments of lower extremities. Promising findings were found as the ANN could reasonably predict the target joint moments. The results also showed the generalisation ability of the ANN to estimate the joint moment that it has not seen before, for instance, the ANN could fairly predict joint moments of the contralateral limb. The Xsens system was validated against the standard motion capture system before the main estimation study of the joint moment in gait began. The results revealed that joint angles obtained from the Xsens were comparable with the optoelectronic system in the sagittal plane and less comparable in the frontal plane according to the coefficient of multiple correlation and the linear fit methods. The results from the transverse plan

关键词： Computer science   Artificial intelligence   Biomedical engineering  

关键词： Energy   Aerospace engineering   Alternative Energy   Atmospheric sciences   Computer science   Engineering   Fluid mechanics   Mechanics   Physics  

摘要： Underwater energy-harvesting kites possess the capacity, under appropriate control algorithms, to deliver more than an order of magnitude more power per unit mass than stationary devices. However, in order to perform path-following control of an underwater kite, it is essential to maintain accurate estimates of the kite’s position and velocity. Obtaining these measurements is complicated by the facts that (i) the kites operate in GPS-denied environments and (ii) tether curvature renders simple line angle sensor-based measurements inaccurate in numerous flow conditions. In this thesis, a closed-loop observer that fuses line angle sensor and accelerometer measurements with a dynamic tether model to maintain a real-time estimate of an underwater kite’s states is presented. The observer design including a robustness to observer tether drag coefficient error was analyzed in simulation. The observer design was experimentally validated on a 1/10-scale kite model, using a customized tow testing system deployed in the NC State Aquatic Center. The observer was validated in an approximately straight tether regime and a curved tether regime. In the straight tether regime, the results were close to a straight-line estimate as expected. In the curved tether regime, the results were compared against underwater photograph. There, the observer estimated position matched the calculated position from the photograph and therefore outperformed the straight-line estimate.

关键词： Neurosciences   Probability   Control algorithms   Brain research   Decision making   Markov analysis   Computer science   Mathematics   Operations research  

关键词： Cold   Software   Stainless steel   Spectrum analysis   Alloys   Physical properties   Lasers   Analytical chemistry   Chemistry   Computer science   Materials science   Optics  

摘要： El objetivo de este trabajo fue analizar el proceso de conformado láser aplicado a dos materiales relevantes para la industria del transporte: Acero IF de alta resistencia y la aleación de aluminio AA 6013. El estudio incorporó trabajo experimental y simulaciones numéricas, considerando dos escenarios distintos de operación, conseguidos al variar la velocidad en el escaneo lineal del haz. Se incorporaron ensayos metalográficos, de GDOES y de difracción de rayos X para la caracterización microestructural. Además, el comportamiento del material durante el proceso se describió usando una formulación termomecánica-plástica. El objetivo final de los resultados numéricos y experimentales es establecer correlaciones entre el ingreso de energía, la microestructura resultante y el ángulo final de conformado. En este espíritu, para poder validar primero los resultados numéricos se realizaron simulaciones de conformado láser en AISI 304, un material que fue adoptado como referencia en este trabajo debido a su extensivo uso en operaciones de plegado láser; mediante la comparación de los resultados numéricos con los correspondientes resultados experimentales y correlacionando además los perfiles de temperatura y microestructura a lo largo del espesor de la placa. Luego, la misma metodología se usó para las muestras de acero IF y AA 6013. Los resultados obtenidos destacan el uso de nuevos enfoques de modelación para tratamientos de conformado láser en aleaciones metálicas avanzadas.