Driver Behaviour Recognition Using Hidden Markov Models

Haghighi Osgouei, Reza Pohang University of Science and Technology, Divis 2012 국내석사

In this work we addressed the problem of modelling human driving behavior using hidden Markov models (HMMs). It is part of a bigger objective towards capturing and transferring driving skills from an expert driver to a novice trainee. We believe driving behaviors are in result of driver's decision making rules. So we drew our attention to identify and recognize driver's decisions or in another sense driving rules using driving time-series signals. For this end, first a driving simulator based on a commercial racing wheel is developed to simulate a desired driving task. The required driving signals including acceleration pedal position, steering wheel angle, velocity and heading of the vehicle are collected using the driving simulator. Then inspired by the fact that the only variables a driver has control on them are velocity and heading, their first-order derivative are extracted as the two most important features of driving patterns. Following the same inspiration, we developed an automatic segmentation method to detect the local extrema of controlling variables and divide data samples into a number of segments. We suggest during each segment, the driver keeps the pedal and wheel operations unchanged. Not all data segments come from different sources; there might be some criteria to group similar segments. In this line we proposed three partitioning methods, threshold-based, GMM-based, and hierarchical, all originated from dividing the two dimensional feature space into a number of classes. In our belief, data classes are the time-domain realization of driving behaviors. According to each class, the parameters of one HMM are optimized to be used later for recognizing driving behaviors. The achieved high average correct classification rate, between 85% to 95% depend on the partitioning criteria, reveals the efficacy of proposed approach in classifying and recognizing driving behaviors. Finally we made use of behavior recognizers to compare an expert and a novice driver's performances in order to provide some feedback. Two methods one road-dependent and another road-independent are proposed for this end. The evaluation results proved the applicability of proposed methods.

Haptic rendering of surface 3D curvature and texture on electrostatic friction display

Haghighi Osgouei, Reza Pohang University of Science and Technology 2018 국내박사

In this work, we address the problem of rendering surface curvature and fine texture using an electrovibration display. We proposed effective algorithms to address each problem separately. In the first part, we introduced a gradient-based method to render 3D objects on an electrovibration display. It includes a generalized real-time algorithm to estimate surface gradient from the surface of any 3D mesh. In addition, a separate edge detection method is included to emphasize sharp edges while scanning the surface of a mesh. Conducting a human user study, we showed that in the presence of haptic feedback generated using our algorithm, the users can better recognize 3D bumps and holes when the visual feedback is limited and puzzling. In the second part, we proposed a neural network based texture modeling and rendering method. We first created an inverse neural network dynamic model for the electrovibration display to estimate an actuation signal from the forces collected from the surface of real texture samples. For the force measurement, we developed a linear motorized tribometer enabling adjusting applied normal pressure and scanning velocity. We trained neural networks to learn from the forces resulted from applying a full-band PRBS (pseudo-random binary signal) to the electrovibration display and generate similar actuation signals. While the networks are trained under known normal pressure and scanning velocity, for untested conditions, we proposed a two-part interpolation scheme to produce actuation signal from the neighborhood conditions. The first part generates a signal by taking a weighted average between the signals with the same scanning velocity but different masses. The second part, performs a re-sampling process, either down-sampling or up-sampling, on the newly estimated signals to produce a final signal according to the user applied normal pressure and scanning velocity. We conducted a user study to put the proposed algorithm to test. We asked users to rate the similarity between a real texture and its virtual counterpart. The experimental setup included a load-cell to measure user applied pressure and an IR-frame to track his/her finger position and eventually calculate user's scanning velocity. Testing six different real texture samples, we achieved an average similarity score of 60% using the proposed algorithm against 39% using a basic record-and-playback method. This revealed the potentials of the proposed texture modeling and rendering algorithm accompanied by a linear interpolation scheme in creating virtual textures similar to the real ones.

Actihaemylbedingte antagonisierung der Allylalkoholschadigung im Elektroretinogramm der Maus

Waladan-Haghighi, Gholam-Hossein Univ. Hamburg 1981 국내박사

Genetic studies of complex diseases: Application and methodology for linkage and parent-of-origin effect analyses

Haghighi, Fatemeh Ghorbanzadeh Columbia University 2000 해외박사(DDOD)

In the work presented here, I investigate the genetic basis of bipolar disorder and panic disorder and present a novel statistical method to test for parent-of-origin effect in complex disorders. I applied a variety of statistical algorithms to elucidate the genetic etiology of these disorders. In the analysis of the bipolar disorder data for chromosome 18, affected sib-pair analyses were carried out and suggestive evidence for linkage to two peri-centromeric markers, D18S45 and D18S53 was observed. In the analysis of the panic disorder data, two putative candidate regions on chromosome 7p15 and 13q32 were identified that may harbor susceptibility genes for the disorder. In both of the genome screens performed, the marker D7S435 gave significant evidence for linkage. Striking evidence for linkage was also obtained at marker D13S779 when the definition of the affected phenotype was extended to include certain medical symptoms, constituting a potential panic syndrome. In addition, a hypothesis-driven candidate gene approach was employed to identify susceptibility loci for panic disorder, including the serotonin transporter (5-HTT), dopamine receptor (DRD4), dopamine transporter (DAT), and Catechol-O-Methyltransferase (COMT). Of these candidate genes, suggestive evidence for linkage or association between panic disorder and polymorphisms in the COMT gene and a nearby marker, D22S944, was observed, indicating the presence of a potential susceptibility locus in the chromosome 22q11 region. I also developed a novel likelihood-based method for testing for parent-of-origin effect in complex diseases. This likelihood-based approach for nuclear families calculates the exact likelihood for all parental mating types, allowing for reduced penetrance. The likelihood model has been extended to incorporate ascertainment as well as differential male and female ascertainment probabilities. My results from computer simulations indicate that consideration of potential ascertainment biases and testing of parent-of-origin effect is important and warrants further investigation. This approach may hold great promise for future research in complex diseases with parental effects.

Controlling Energy-Efficient Buildings in the Context of Smart Grid: A Cyber Physical System Approach

Haghighi, Mehdi Maasoumy University of California, Berkeley 2013 해외박사(DDOD)

The building sector is responsible for about 40% of energy consumption, 40% of greenhouse gas emissions, and 70% of electricity use in the US. Over 50% of the energy consumed in buildings is directly related to space heating, cooling and ventilation. Optimal control of heating, ventilation and air conditioning (HVAC) systems is crucial for reducing energy consumption in buildings. We present a physics-based mathematical model of thermal behavior of buildings, along with a novel Parameter Adaptive Building (PAB) model framework to update the model parameters, as new measurements arrive, to reduce the model uncertainties. We then present a Model Predictive Control (MPC), and a Robust Model Predictive Control (RMPC) algorithm and a methodology for selecting a controller type, i.e. RMPC or MPC, versus Rule Based Control (RBC) as a function of model uncertainty. We then address the Cyber-Physical" aspect of a building HVAC system in the design flow. We present a co-design framework that analyzes the interaction between the control algorithm and the embedded platform through a set of interface variables, and demonstrate how the design space is explored to optimize the energy cost and monetary cost, while satisfying the constraints for occupant comfort level. The last part of this dissertation is centered on the role of smart buildings in the context of the smart grid. Commercial buildings have inherent flexibility in how their HVAC systems consume electricity. We first propose a means to define and quantify the flexibility of a commercial building. We then present a contractual framework that could be used by the building operator and the utility company to declare flexibility on one side and reward structure on the other side. We also present a control mechanism for the building to decide its flexibility for the next contractual period to maximize the reward. We also present a Model Predictive Control (MPC) scheme to direct the ancillary service power flow from buildings to improve upon the classical Automatic Generation Control (AGC) practice. We show how constraints such as slow and fast ramping rates for various ancillary service providers, and short-term load forecast information can be integrated into the proposed MPC framework. Finally, results from at-scale experiments are presented to demonstrate the feasibility of the proposed algorithm.

Spatio-Temporal Logics for Verification and Control of Networked Systems

Haghighi, Iman Boston University ProQuest Dissertations & Theses 2019 해외박사(DDOD)

Emergent behaviors in networks of locally interacting dynamical systems have been a topic of great interest in recent years. As the complexity of these systems increases, so does the range of emergent properties that they exhibit. Due to recent developments in areas such as synthetic biology and multi-agent robotics, there has been a growing necessity for a formal and automated framework for studying global behaviors in such networks. We propose a formal methods approach for describing, verifying, and synthesizing complex spatial and temporal network properties.Two novel logics are introduced in the first part of this dissertation: Tree Spatial Superposition Logic (TSSL) and Spatial Temporal Logic (SpaTeL). The former is a purely spatial logic capable of formally describing global spatial patterns. The latter is a temporal extension of TSSL and is ideal for expressing how patterns evolve over time. We demonstrate how machine learning techniques can be utilized to learn logical descriptors from labeled and unlabeled system outputs. Moreover, these logics are equipped with quantitative semantics and thus provide a metric for distance to satisfaction for randomly generated system trajectories. We illustrate how this metric is used in a statistical model checking framework for verification of networks of stochastic systems.The parameter synthesis problem is considered in the second part, where the goal is to determine static system parameters that lead to the emergence of desired global behaviors. We use quantitative semantics to formulate optimization procedures with the purpose of tuning system inputs. Particle swarm optimization is employed to efficiently solve these optimization problems, and the efficacy of this framework is demonstrated in two applications: biological cell networks and smart power grids.The focus of the third part is the control synthesis problem, where the objective is to find time-varying control strategies. We propose two approaches to solve this problem: an exact solution based on mixed integer linear programming, and an approximate solution based on gradient descent. These algorithms are not restricted to the logics introduced in this dissertation and can be applied to other existing logics in the literature. Finally, the capabilities of our framework are shown in the context of multi-agent robotics and robotic swarms.

Stochastic load models from limited data: A general approach with applications to wind and waves

Kashef Haghighi, Tina Stanford University 2002 해외박사(DDOD)

Over the last few decades the design of structures has become increasingly based on a probabilistic description of the variables involved. Among these variables, the loads applied to the structure are perhaps the most critical and difficult to model, due to their inherent randomness and the difficulty of obtaining loads data in extreme conditions. Here we present a review of stochastic load modeling, followed by our work on three different topics in this field. We focus on developing models that take into account the limited nature of available data. We first consider the problem of finding the short-term load distributions within a particular environment condition. We use a moment-based approach to fit parametric probability distributions to the loads. We show that one-sided distributions that match four statistical moments estimated from limited loads data may show erratic behavior in the distribution tail. In particular, for modeling fatigue loads on wind-turbine blades, the quadratic Weibull distribution is found to be more robust than the cubic Weibull distribution suggested previously. We then use a regression model to estimate the moments of the short-term distributions from environment parameters and suggest a method to evaluate the importance of different environment parameters in predicting the load moments. We apply this approach to the problem of modeling fatigue loads on wind-turbine blades and show that for two of the three turbines studied, a parameter found by high-pass filtering the wind speed can explain a large fraction of the variations in the load moments. For a third turbine, the mean wind speed is found to explain an even larger fraction of the moment variations, suggesting that the important wind parameters may be structure-dependent. Finally, we discuss the calibration of response prediction models. Calibration methods based on marginal error statistics or bias-adjustment factors, inherent in many structural design codes, are shown to be insufficient for predicting the distribution of measured loads. A linear regression model is shown to solve this problem and is used to calibrate fluid drag load models for offshore structures and to examine the effect of various model parameters on the accuracy of the response predictions.