Abstract

Estimation of time delay from the received broadband signals like speech, collected at two or more spatially distributed microphones, has many applications, ranging from speaker localization, speaker separation, determining the number of speakers and tracking a moving speaker. Cross-correlation based methods like the cross-correlation of signals directly and generalized cross-correlation based methods (GCC, GCC-PHAT) have been used for several years to estimate the time delay. Performance of these methods degrades due to noise, multipath reflections, and reverberation in a practical environment, like a live room. The objective of this thesis is to address the issues of robustness and accuracy of the estimated time delay by generating multiple evidence from several cross-correlation sequences using single frequency filtering (SFF) method for decomposition of the received signals into components, one at each frequency. The accuracy of the estimated delay is improved by exploiting the amplitude information of the samples of the cross-correlation sequence around the peak corresponding to the integer delay. The estimated fractional part of the delay is shown to be not only accurate, but also robust against different types and levels of additive noise degradations. The accuracy of the estimation of the fractional part is further improved by using a trained artificial neural network model for regression. Simulation experiments with known time delay are used to evaluate the accuracy and robustness of the proposed time delay estimation methods. Complementary information to the SFF magnitude is provided by the phase component of the SFF output. The time delays are estimated accurately and independently by using the signals reconstructed from the SFF phase alone. The complex SFF spectra at each instant are used to obtain the instantaneous time delay. The instantaneous time delay is exploited for separation of individual speakers from the mixed signals received at two spatially separated microphones in a live room. The proposed time delay estimation method is also used for determining the number of speakers from the mixed signals. Applications of the time delay estimation methods are illustrated using the actual field speech data of speakers, collected at two spatially separated microphones in a live laboratory room, while the speakers are reading a text simultaneously. Performance of the proposed time delay estimation methods is evaluated in terms of accuracy and robustness.

Abstract

In this thesis, we develop a higher understanding of a dynamic road scene by Leveraging the information present in the evolution of spatial relations present between the objects of interest in the road scene. We first propose a novel motion clustering formulation over Spatio-temporal depth images obtained from stereo sequences that segments multiple motion models in the scene in an unsupervised manner. The motion models are obtained at frame rates that compete with the speed of the stereo depth computation. This is possible due to a decoupling framework that first delineates spatial clusters and subsequently assigns motion labels to each of these clusters with analysis of a novel motion graph model. A principled computation of the weights of the motion graph that signifies the relative shear and stretch between possible clusters lends itself to a high fidelity segmentation of the motion models in the scene. The fidelity is vindicated through accuracies reaching 89.61% on KITTI and complex native sequences. In the second part of thesis, we develop a semantically meaningful awareness of the dynamic road scenes for Understanding the on-road behavior of vehicles Humans typically can decompose a dynamic scene as vehicles exhibiting behaviors like ”moving away from me”, ”moving towards me”, ”changing lane” and the like. Such higher-level scene awareness can have implications on the performance of algorithms lower in the hierarchy. For example, a SLAM or state estimation algorithm can benefit from the knowledge that a car is ”parked” as features from a parked car is indeed useful for estimating relative camera motion. In this paper we decompose a road scene into vehicle behaviors such as ”parked”, ”following lane - moving away”, ”following lane - coming towards” and similar such labels. We accomplish this through Multi-Relational Graph Convolutional Networks (MR-GCN) that is proving itself to be an apt architecture to learn behaviors of agents or objects based on their evolving relations with other agents or objects in the scene. We show high fidelity behavior prediction of vehicles encountered in a variety of datasets such as Cityscapes, KITTI, ApolloScape, and Indian. More critically we show the effective model transfer from one dataset to another verifying generality and repeatability across a diverse and rich set of worldwide scenes.

Abstract

In recent years there has been a growing interest in understanding natural language for the purpose of answering questions related to science and maths. Automatically generating computer programs to solve problems posed in natural language is one of the holy grail problems of AI. A step by step approach to this would first involve understanding the problem from the text and then categorising this problem into certain categories based on algorithms and many other criteria and then finally generating the code. In this thesis we tackle the problem of automatically categorising mathematical and programming word problems. Recent advances in Deep Learning, Machine Learning based classifiers as well as the advent of word embeddings has shown impressive results in the field of Text Classification. In light of the advances in text classification we model the task of operator prediction in math word problems and algorithm prediction in programming word problems as a text classification problem and solve it using the state of the art approaches for text classification. In the first part of this thesis we deal with arithmetic word problems. Arithmetic word problems can be solved with the help of the numbers mentioned in the text and their relationships through basic mathematical operations (addition, subtraction, division, multiplication). We start by creating a model which detects the mathematical operation (+, -, *, /) that would be used to solve an arithmetic word problem. We then use this model to build an arithmetic word problem solver, a solver which takes as an input an arithmetic word problem and outputs the answer. We use various machine learning and deep learning based algorithms to build the model which detects the mathematical operation. Our solver is able to solve 81 percent of all the problems correctly in our dataset. We then broaden our focus to deal with programming word problems. These are problems in natural language which can be solved by a computer program. These problems are significantly harder than math problems for the following reasons. The solver must first decode the intent of the problem, or understand what the problem is. Then the solver needs to apply their knowledge of algorithms to write a solution program. Another reason is that, the solution programs must be efficient with respect to the given time and memory constraints. An outgrowth of this is that, the algorithm required to solve a particular problem not only depends on the problem statement, but also the constraints. These problems are routinely used by major tech companies to hire software engineers. Hence, these problems represent a significant problem for AI which has largely been untouched in current literature. As a first step towards solving these problems we introduce the task of algorithm prediction for programming word problems. We formulate the problem as a single and multi-label classification task and use state of the art Machine Learning and Deep Learning Based classification algorithms on this task. We create the first publicly available dataset for the task of algorithm prediction in a programming problem. From this dataset, we create 4 classification tasks; 2 single label classification tasks with 5, 10 classes respectively and 2 multi-label classification tasks with 10, 20 classes respectively. Our best performing classifier gets an accuracy of 62.7 percent on the 5 class single label classification whereas the best multilabel classification model has an accuracy only 9 percent lower than a skilled human. To the best of our knowledge, these are the first reported results on such a task. We make our code and datasets publicly available.

Abstract

Designing communication systems and protocols in the real-world often involves various subproblems such as: a) “what information should be sent?”, b)“how to syntactically encode this information to messages?”, c) “how to reliably transmit these messages?” among other similar questions, and d) since communication is often means to some end, “how can the communicated information be used at destination?”. While information theory and communication engineering allow us to approach a significant half of these problems, aspects such as ”what should be communicated?” and ”how should the information be used?” often demand accounting for the context and underlying objectives of designing communication systems. Building upon economic theories of communication and signaling, we approach the problem of engineering of communicative structures within an economic paradigm, and discuss various algorithmic and game-theoretic issues that emerge from this exploration. We first study the problem of optimal lossy compressions – of comparing different ways to compress a piece of information on the basis of economic utility – when communication takes place between cooperating agents. Using ideas from combinatorial optimization, we present approximation algorithms for the simple version of the problem, as well as in the presence of side-information and in presence of uniform noise. We then study the problem of designing communicative structures in presence of strategic behaviour from Sender and Receiver. We discuss three forms of design problems in this model: a) Sender-led design, b) Receiver-led design, and c) Design by some third-party principal. For the first two, we discuss the general model, Best Response characterizations, Stackelberg (or leader-follower) equilibriums, and the relevant versions of revelation principle. We also present the associated linear programs for solving them under unbounded communication and noiseless conditions, and discuss the relationship with related ideas in Persuasion and Mechanism Design. We also show the existence of instances of Sender-led (or Receiver-led) design problem where the presence of exogenous (stochastic) noise has gainful (or welfare-positive) effects for the Receiver (or the Sender respectively).Further, for the problem of design by third-party principals, we propose a novel methodology of empowering the third-party designer (say, Wirepuller) with ability to deliberately design channel noise. We discuss how this simplifies the problem of selecting some optimal protocol such that neither the Sender not Receiver is incentivized to deviate from it, and provide the two-sided relevation principle for the problem, and present the corresponding linear programming formulation. We also discuss how this framework provides an economic model of ambiguous signalling conventions. Finally, we provide a preliminary discussion of the sequential version of this model, with infinitehorizon discounted rewards as objectives, in terms of communicating states of Markov Decision Processes (MDPs). This allows us to move beyond the static assumptions, towards designing continual communication systems with long-term preferences. We provide an analysis and demonstrate the sufficiency of single state history for the Receiver-led design, i.e. policy that is strategyproof (or incentive compatible) for planning under strategic revelation of state information.

Abstract

Increasing the spectral efficiency of the cellular networks has been and continues to be one of the major factors from 1G to 5G and beyond. In this regard, cellular operators have turned towards heterogeneous cellular networks (HetNets) to improve the area spectral efficiency of the networks. One of the significant challenges in HetNets is the underlying co-channel interference that is experienced by the users. Though the network throughput increases due to a better area spectral efficiency of a HetNet, there is a possibility that high interference will make few link capacities close to zero when users regard interference as noise (IAN). In other words, the dynamic nature of these co-channel interference links is also a problem due to its threat of causing outage or reduction in quality of service (QoS) to the UEs. Instead of trying to avoid interference, nodes in the network can also exploit it if they are equipped with such capabilities. In this work, successive interference cancellation (SIC) technique is investigated to reduce the crosstier interference in a reverse time division duplexing (RTDD) two-tier HetNet. In this work, the sum link capacity expressions for an RTDD HetNet, where the users use SIC to remove the cross-tier interference is derived. The derived sum link capacity expression for SIC is compared with that of IAN to derive a switching condition on using SIC. This opportunistic use of SIC (which is referred to as switching in this thesis) ensures a minimum sum link capacity (proportional to the desired signal strengths) independent of the interference links, thereby solving the QoS problem. A network operator can also remove the interference dependency by splitting the available resources between the tiers. However, we have proven that the maximum sum link capacity that can be achieved by orthogonal resource allocation schemes is the same as the minimum that is ensured by switching. System-level simulations are done to show the gain in the overall system capacity of switching compared to only using either IAN or SIC for randomly paired UEs. The system capacity for IAN, SIC, and switching can be further improved by optimizing the selection of UEs across the tier (user pairing) that use the same frequency resources. In this regard, the user pairing scheme that should be used for IAN, SIC, and switching cases is formulated as a discrete optimization problem. This optimization problem also encompasses the cases where all the UEs do not need to be paired. The Hungarian algorithm is used to solve this optimization problem after some prepossessing done to the cost matrix. Through system-level simulations, the gains in system capacities for Hungarian paired UE using IAN, SIC, and switching is shown compared to their random counterparts.The above-discussed results assume that the UE has the necessary demand for the supported downlink (DL) and uplink (UL) data rates. In the case of any traffic asymmetries, the subframe allocation can be dynamically adapted in a synchronized TDD HetNet. However, the inability of a synchronized RTDD HetNet’s frame structure questions the usefulness of these results when there is an asymmetric data demand (DLUL). In such cases, the UE does not need to operate at its full capacity in UL. Hence, it would throttle either the transmission rate or power. Rate expressions for rate and power throttling are derived for cases where UE uses SIC or regards IAN. System-level simulations are performed to compare the overall system throughput of IAN and SIC for different traffic asymmetries (DL:UL ratios). It is shown that if UEs does rate throttling, then using SIC would improve the overall achievable rates compared to IAN, and vice-versa for power throttling

Abstract

Water quality assessment has undeniable value for humanity’s well-being. It is a principle strategy that aids in ascertaining water contaminants, providing information about the state of water supplies, and plays a vital role in their efficient management. Despite that, most of the existing techniques struggle to rapidly measure the quality of water due to costly infrastructure or large evaluation time. Recently, quality measurement has been demonstrated using spectrum analysis. Although this method has been investigated in studies, the focus has been on analysing spectrum of wavelengths to predict the quality of water, which relies on commercial spectrometers for spectrum acquisition. However, these modern systems need complicated signal processing algorithms and high processor capabilities for decision making. They have elaborate optical assemblies and sizable power requirements, making them bulky and expensive to operate. Thus, a refined technique with simplified implementation, low cost, and reliable long-time operation is essential to address this challenge. We present a simplified approach towards a rapid method for determining water quality parameters in sampled and flowing water. This is based on the selection of characteristic wavelengths for the parameters considered and the usage of narrowband LEDs as the inspection light source. Thereby incorporating the benefits of optical sensing such as electromagnetic immunity, selectivity, sensitivity, etc. The specific wavelengths of 560 nm, 860 nm, and 635 nm have been demonstrated to have a dominant effect due to pH, turbidity, and total dissolved solids (TDS), respectively, from the regression analysis. Partial least square regression was applied on the dataset of absorption spectra prepared to estimate the significant wavelengths. This selection of wavelengths was supported by determining wavelength significance using neighbourhood components analysis, which gave similar wavelengths. To visualise the spread and distribution in spectral data, t-distributed stochastic neighbour embedding was used. Using only these wavelengths, an evaluation system capable of determining the light absorption after passing through water has been designed and developed. A prototype structure was designed, and LEDs of the selected wavelengths were utilized as light sources along with the corresponding photodiode to determine absorbance. The obtained optical responses are subsequently related to water parameters, specifically pH, turbidity, and TDS. Experiments were performed to evaluate samples and then validate this technique against standard instruments for flowing and sampled water setups. It is shown from the measurement results that pH, turbidity, and TDS have linear regression coefficients of 0.9773, 0.9617, 0.8271 and 0.9691, 0.9729, 0.76 for flowing and sampled water arrangements, respectively. This study asserts the application of wavelength-specific absorption to assess water quality parameters and provides results based on spectral analysis and experimental implementation using LEDs, leading to a similar conclusion.

Abstract

At present, 3.6 billion people (nearly half of the world population) live in areas under high water scarcity according to World Water Development (UN) report. It is expected that more than 70% of the global population will live in urban areas by 2050. Contiguously, the water crisis is a major issue across various cities in the world due to the shrinking of water resources because of climate change, global warming, etc. Hence, it becomes arduous for the water authority to feed and provide potable water to such a large population. Using sensors, ICT (Information and communication technology) and BDA (Big Data Analytics) water resources can be managed and saved for future use. Thus, there is an exigency for smart devices which can provide real-time water usage data. Today, the privatization of water metering sector is leading to the surge of cost for smart water meter in urban areas. This thesis offers a solution to the problem in two ways: First by designing an affordable smart water meter using available flow sensors in market and other by suggesting a noninvasive approach to counter the issues with the smart water meter having mechanical flow sensor. Here, an algorithm for calibration of available flow sensors is explained by designing an experimental setup. The experimental results show that the given set up can be used to calibrate various sizes of a flow sensor with an accuracy of ±0.17% to ±1.22%. Then two architectures using GSM and Wi-Fi protocols are developed for processing the data available from the flow sensor and transmitting to the cloud using the IoT features. Finally, a 3D design is suggested for packaging of a calibrated flow sensor and proposed architecture designed at previous stages to form a complete smart water meter. The experimental and simulation results show that the proposed architecture for smart water meter has characteristics of low power, suitable area and affordable cost. An alternative flow sensing technique is proposed and implemented to eradicate the constraints of mechanical flow sensor. An ultrasonic signal based clamp-on type flow sensor is presented. The experimental results show that the suggested technique can be useful in designing of water pipeline monitoring system

Abstract

Monocular 3D human reconstruction is a very relevant problem due to numerous applications to the entertainment industry, e-commerce, health care, mobile-based AR/VR platforms, etc. However, it is severely ill-posed due to self-occlusions from complex body poses and shapes, clothing obstructions, lack of surface texture, background clutter, single view, etc. Conventional approaches address these challenges by using different sensing systems - marker-based, marker-less multi-view cameras, inertial sensors, and 3D scanners. Although effective, such methods are often expensive and have limited wide-scale applicability. In an attempt to produce scalable solutions, a few have focused on fitting statistical body models to monocular images, but are susceptible to the costly optimization process. Recent efforts focus on using data-driven algorithms such as deep learning to learn priors directly from data. However, they focus on template model recovery, rigid object reconstruction, or propose paradigms that don’t directly extend to recovering personalized models. To predict accurate surface geometry, our first attempt was VolumeNet, which predicted a 3D occupancy grid from a monocular image. This was the first of its kind model for non-rigid human shapes at that time. To circumvent the ill-posed nature of this problem (aggravated by an unbounded 3D representation), we follow the ideology of providing maximal training priors with our unique training paradigms, to enable testing with minimal information. As we did not impose any body-model based constraint, we were able to recover deformations induced by free-form clothing. Further, we extended VolumeNet to PoShNet by decoupling Pose and Shape, in which we learn the volumetric pose first, and use it as a prior for learning the volumetric shape, thereby recovering a more accurate surface. Although volumetric regression enables recovering a more accurate surface reconstruction, they do so without an animatable skeleton. Further, such methods yield reconstructions of low resolution at higher computational cost (regression over the cubic voxel grid) and often suffer from an inconsistent topology via broken or partial body parts. Hence, statistical body models become a natural choice to offset the ill-posed nature of this problem. Although theoretically, they are low dimensional, learning such models has been challenging due to the complex non-linear mapping from the image to the relative axis-angle representation. Hence, most solutions rely on different projections of the underlying mesh (2D/3D keypoints, silhouettes, etc.). To simplify the learning process, we propose the CR framework that uses classification as a prior for guiding the regression’s learning process. Although recovering personalized models with high-resolution meshes isn’t a possibility in this space, the framework shows that learning such template models can be difficult without additional supervision. As an alternative to directly learning parametric models, we propose HumanMeshNet to learn an “implicitly structured point cloud”, in which we make use of the mesh topology as a prior to enable better learning. We hypothesize that instead of learning the highly non-linear SMPL parameters, learning its corresponding point cloud (although high dimensional) and enforcing the same parametric template topology on it is an easier task. This proposed paradigm can theoretically learn local surface deformations that the body model based PCA space can’t capture. Further, going ahead, attempting to produce highresolution meshes (with accurate geometry details) is a natural extension that is easier in 3D space than in the parametric one. In summary, in this thesis, we attempt to address several of the aforementioned challenges and empower machines with the capability to interpret a 3D human body model (pose and shape) from a single image in a manner that is non-intrusive, inexpensive and scalable. In doing so, we explore different 3D representations that are capable of producing accurate surface geometry, with a long-term goal of recovering personalized 3D human models.

Abstract

Cuff-less blood pressure (BP) is essential for continuous health monitoring to prevent diseases such as hypertension. Due to the discomfort caused by inflation and deflation of the cuff, it is not possible to monitor continuously. Although Pulse Arrival Time (PAT) derived from Electrocardiogram (ECG) and Photoplethysmogram (PPG) for cuff-less Blood Pressure (BP) measurement has been a contemporary and widely accepted technique. However, the features extracted for it are conventionally from an isolated pulse of ECG and PPG signals. As a result, the estimated BP is intermittent. Multi-parameter models are developed using regression analysis, to estimate systolic blood pressure (SBP) and diastolic blood pressure (DBP). Hence, the correlation of multiple extracted features with the blood pressure is proved. To achieve this, simultaneous electrocardiogram (ECG) and photoplethysmographic (PPG) along with respective BP data were collected for 171 patients as per Association for the Advancement of Medical Instrumentation (AAMI) standards from Care Hospital, Hyderabad. During data collection, the simultaneous ECG and PPG data was collected from (Food and Drug Administration) FDA approved Vios Medical System (VMS) and BP data was collected using mercury based sphygmomanometer, manually after the required training. The developed algorithm uses wavelet transformation on ECG and PPG signals for detection of the occurrence of essential wave points precisely, even in the presence of artifacts. The feature vector of dimension 32, has been created using the fiducial points detected from ECG and PPG signals. The SBP and DBP estimation model has been obtained using regression analysis. The proposed estimation model provides the error within the range proposed within AAMI and British Hypertension Society(BHS) standards for BP estimation. In addition to this, Central BP (CBP), which is more closely associated with the cardiac activities, has also been estimated. Data for 33 patients was collected from Cath lab and CBP estimation model has been trained using PLSR regression analysis. As a resultant, fair amount of estimation accuracy has been achieved. Further, hardware design has been proposed which will simultaneously measure ECG and PPG signal and transmit the data to processing unit via bluetooth. The data received at processing unit will be processed and peripheral BP and CBP are estimated using the proposed algorithms. This continuous real-time BP monitoring technique can be useful in the treatment of hypertensive and potential-hypertensive subjects

Abstract

The pervasive nature of music in human lives, from day to day activities to large scale concerts makes it an important part of human social interactions. From an evolutionary perspective, it has been argued that music has the capacity to evoke shared responses within groups, thereby allowing it to act as a form of social glue increasing group cohesion. One proposed mechanism for these shared responses is empathy. Over the last decade, a great deal of research has sought to identify the behavioural mechanisms underlying the affective capacity that music affords. These studies have put forth various theories that suggest an integral role played by empathy in inducing emotions while listening to music. In this thesis, we try to expand our understanding of empathy in the context of music processing from both behavioural and neural fronts. Through an extensive literature review, we find that previous studies on empathy have sought to examine it in terms of two distinct but related subsystems: The cognitive subsystem, which is concerned with top-down aspects of empathy such as motor-mimicry and the affective subsystem, which primarily deals with bottom-up aspects related to affect processing such as emotional contagion. In the first study, we looked at behavioural differences in self-reported continuous ratings of ”feeling moved” while participants listened to continuous pieces of music. Using a data-driven approach to cluster the ratings, we found significant differences in ”feeling moved” to both sad and joyful music modulated by participants’ affective empathy. Our results provided empirical support to the aforementioned Cognitive vs. Affective framework of empathy from a behavioural standpoint. In the second study, we sought to extend our earlier work by examining how empathy modulated brain responses while listening to music using Functional Magnetic Resonance Imaging (fMRI). Prior to the current study, neuroimaging studies investigating the role of empathy in music processing have been scarce and did not examine functional connectivity in the brain. Our study overcomes this limitation by examining both global and localised processing of music modulated by empathy through graph-theory based methods. Overall, we found differences in global and local connectivity patterns in music processing modulated by participants’ trait empathy. Participants scoring higher on cognitive empathy tended to recruit mechanisms related to motor mimicry, whereas those scoring higher on affective empathy were found to recruit regions related to auditory affect processing. Both of our studies corroborate previous research in the field in addition to providing novel findings and as a result further supporting the existence of two subsystems in Cognitive vs. Affective framework of understanding empathy.