Research Advisor : Prof. Yogesh Simmhan
Research Topic : Systems for Scalable City-Wide Video Analytics
Abstract: Advances in deep neural networks (DNN) and computer vision (CV) algorithms have made it feasible to extract meaningful insights from large-scale deployments of urban cameras. However, it is impractical to execute the full video analytics pipeline on all the cameras due to the punitive computing and network costs. Smart city video analytics can be deployed across edge, fog, and cloud resources with accelerators onboard, and may be mobile. This creates several challenges such as intermittent network disconnections, and spikes in input data. I explore the System design and Programming abstractions for such analytics applications.
Research Advisor : Prof. Radhkant Padhi
Research Topic : Optimal and Safety Critical Control for Artificial Pancreas System
Abstract: The aim of this research is to develop control techniques for safety critical cyber-physical systems. Safety is an important aspect of control systems where bounds/limits need to be respected. The application that this research will be focusing on is the development of Artificial Pancreas (AP) system for Type 1 Diabetes Mellitus patients in India. T1DM patients do not produce insulin in their body to regulate blood glucose. This condition can cause serious health risks due to hypo and hyper glycemia. This can be prevented by providing insulin externally through an insulin pump. Most commercially available insulin pumps must be programmed manually and requires constant user intervention and is prone to error. The AP system aims to automate the delivery of insulin by constantly monitoring the glucose in the blood using a Continuous Glucose Monitor. Biomedical systems such as Artificial Pancreas that are used for monitoring and controlling blood glucose are extremely safety critical. Methods from control theory that respect safety bounds (such as model predictive control (MPC)) and estimation methods are key modules in this system. Once basic functionality of the AP is proven, more sophisticated techniques such as online adaptive methods will be incorporated into the system. This will enable the system to learn the patient model in real time. Towards the end of the research, a completely autonomous Artificial Pancreas system which utilizes methods and techniques from optimal control, estimation and A.I. will be developed. Once the performance of this system is proven, it can significantly improve the lifestyle of T1DM patients in India.
Research Advisor : Prof. Pavankumar Tallapragada
Research Topic : Resource Aware Event-Triggered Control of Networked Control Systems
Abstract: A networked control system has different field of applications, such as industrial automation, environment monitoring, healthcare applications, military surveillance and disaster management. One of the main challenges in networked control systems are resource constraints, especially computation, communication, and energy resource constraints. These constraints obstruct the use of classical control techniques and call for new resource-aware control paradigms. Event-triggered control has been an active area of research in the last decade. The main advantage of event triggered control is the efficient aperiodic state dependent sampling of the control while simultaneously achieving control objectives. One of the factors which plays an important role in balancing control objectives with efficient use of resources in event-triggered control is the inter-event time. Analysis of the evolution of inter-event times helps to schedule multiple processes over a shared communication channel or to plan transmissions under constraints.
In my current work a systematic analysis of the evolution of inter-event times, for planar linear systems under a general class of scale-invariant event-triggering rules, is carried out. For scale-invariant event triggering rules, the inter-event time is a function of only the “angle” of the state at an event. The properties of this inter-event time function, such as periodicity and continuity, are analyzed. In particular, the interevent time function is continuous except for finitely many “angles” and sufficient conditions, under which the inter-event time function is continuous, are provided. Then, the evolution of the “angle” of the state from one event to the next is analyzed and the problem of studying the evolution of the inter-event times is reduced to that of studying the “angle” map and its fixed points. For a specific triggering rule, necessary conditions are provided for the convergence of inter-event times to a steady state value. The proposed results are illustrated through numerical simulations.
Future work includes analysis of the angle map under specific triggering rules with regard to necessary and sufficient conditions for the existence of fixed points, their stability, region of convergence and rates of convergence. Extensions to periodic event-triggered control or region based self-triggered control are other avenues for future work. Further research directions include resource aware control of networked control systems with energy or communication constraints and privacy issues involved in such systems. Online co-design of controller and event triggering mechanism is also another direction of research.
Girija Ramesan Karthik
Research Advisor : Prof. Prasanta Kumar Ghosh
Research Topic : Algorithms for low-cost, portable tomographic devices with applications in biomedical imaging and human computer interaction
Abstract: Some of the most important tools in medical diagnostics are imaging devices like the X-ray, CT, MRI etc. They help in visualizing the internal organs of patients, thereby aiding in the diagnosis of the underlying pathology. However, these devices are not portable and incur huge maintenance costs which in turn lead to their low accessibility especially in remote areas and high demand-supply ratio. They are also not-suitable for continuous monitoring of patient parameters. Ultrasound is reasonably portable and can be used for continuous monitoring of the patients. However, when it comes to lung disease diagnosis, conventional ultrasound cannot penetrate the lungs due to high reflections at the air-tissue interface. To overcome these difficulties, among others, researchers have been exploring low-cost and portable imaging modalities like microwave imaging, Electrical Impedance Tomography (EIT), low-frequency ultrasound etc. Recent studies have shown that unlike conventional ultrasound, low frequency ultrasound does penetrate the air-tissue interface and can be used to image the lungs. Microwave imaging has been used for the detection of breast cancer, brain stroke and lung fluid accumulation. Similarly, EIT has been introduced in clinical settings for continuous lung ventilation imaging. In addition to medical imaging, the most interesting aspect of these imaging modalities is that they can be extended to other applications among which of particular interest is human-computer interaction (HCI) (eg. interactive surfaces and hand gesture-recognition). Motivated by these developments, our goal is to develop and improve upon the algorithms for these imaging modalities. The most important problem with these imaging modalities is that the inverse problem is non-linear and ill-posed. The non-linearity implies that most algorithms are iterative and the ill-posedness can cause implausible reconstructions. Regularization has been the go-to method for tackling ill-posedness. However, conventional regularization methods do not incorporate domain specific knowledge about the structure and properties of the internal organs. In addition, the iterative nature of the algorithms make them unsuitable for real-time imaging. With the advent of Deep Neural Networks (DNNs) and the availability of large databases of high resolution CT and MRI scans, domain specific spatial priors can be learnt to provide better regularization and reconstruction. DNNs can also be used to replace the iterative algorithms, hence making the reconstruction process much faster and suitable for real-time imaging. In this work we would like to address three major aspects. Firstly, DNNs trained on a particular imaging configuration (eg. transceiver locations, frequency etc.) when used in other configurations can result in performance degradation. So, we would like to develop methods for light-weight adaptation of these networks to arbitrary imaging configurations. Secondly, we would like to develop methods to extract spatial priors from the databases of CT and MRI that can in turn be
incorporated into the reconstruction process. Thirdly, we would like to extend the conventional reconstruction algorithms to track dynamic organs. Inspired by a recent work that uses EIT for static hand-gesture recognition by estimating the tomographic cross section of the arm, we would like to perform experiments using an EIT setup to extend static gesture recognition to dynamic hand movement tracking. We propose to learn the temporal dynamics using the data collected from a multi-modal setup that can measure EIT data from the arm along with the corresponding hand movements using a
motion-capture setup such as Electromagnetic Articulograph (EMA) or Optitrack.
Lima Agnel Tony
Research Advisor : Prof. Debasish Ghose
Research Topic : Collision avoidance of UAVs with applications to regulated drone traffic
Abstract: Past decade has witnessed a significant growth in the research on unmanned systems and effective utilization of airspace. Considering the numerous applications of Unmanned Aerial Vehicles (UAVs), the relevance of developing practical solutions are more than ever. Similar to the road traffic system which has reached its maturity over the years, the unmanned aerial transport also requires a traffic management system, as more safety critical aspects need to be addressed in this than the former. Collision avoidance is of prime importance in this regard. The uncertainties involved and the limitations in the current technologies demand an efficient avoidance subsystem. This work focuses on avoidance strategies of UAVs while integrating them into a regulated aerial traffic and formally define the underlying rules that streamline the system safety. It also looks in to the development of an implementable solution to a cloud based aerial traffic handling system with minimal supervision and infrastructure/software requirements and assured safety.
Research Advisor : Prof. Parimal Parag
Research Topic : Minimizing latency in data acquisition and distributed processing for cyberphysical systems
Abstract: Cyber-physical systems are an integration of physical processes and inter-connected communication and control systems. Generally, in such systems, a distributed network of sensors record the data regarding the physical process, that is processed and transmitted to an intelligent remote system that responds to the observations, and send the control signal to the actuators. In such applications, the timeliness of data reception and processing is of utmost importance. The research aims at studying latency optimal techniques to do surveillance and processing of data for applications in cyberphysical systems.
Research Advisor : Prof. Chandrasekhar Seelamantula
Research Topic : Theoretical Analysis of Generative Adversarial Networks (GANs), with Applications to Autonomous Vehicle (AV) Tasks
Abstract: GANs are an elegant framework for training generative models to learn a target data distribution, and are a two-player game, with a generator that creates samples that mimic the given training data, while a discriminator learns to differentiate between the real samples and the fake ones. Since their conception in 2014, GANs have found immense success in generating realistic images, image super-resolution, and cross-domain image translation from simulation to real-world environments. While existing literature relies heavily on the empirical success of GANs, the central focus of this Ph.D. is to develop a strong foundational analysis of GANs within a functional optimization framework, drawing parallels to classical signal processing techniques, and leveraging these results to improve the performance of state-of-the-art GAN based AV applications.
Vinay Krishna Sharma
Research Advisor : Prof. Pradipta Biswas
Research Topic : Multimodal Human Robot Interaction
Abstract: The use and scope of robots is increasing day by day as technology advances. Robots are crossing the industrial boundaries they were once confined to, and entering our day to day life as picking and dropping packages in warehouses, serving orders in restaurants, giving company and care to elderly, delivering medicines and food to contagious patients, cleaning and disinfecting hospitals in the current scenario of Covid-19. As these robots finds their place in normal workplaces around us, the interaction between humans and robots becomes unavoidable. People still have some inhibitions, fears, or insecurities about working with or around robots mainly because of our different cultural backgrounds, knowledge, past experiences, and exposure to technology. There is a need of having a simple, natural, human-like, and intuitive interaction between humans and robots. Human Robot Interaction (HRI) is a field of study dedicated to understanding, designing, and evaluating robotic systems for use by or with humans. This research focuses on investigating multi-modal human robot interaction involving non-traditional modalities of interaction like eye-gaze, hand-body gesture, speech and so on. The goals of this research include designing and developing intelligent multimodal HRI systems to work in collaboration with humans. This research aims to extend human interaction with machines (robots) to improve safety and reduce human effort by sharing low level (dirty, dangerous, boring repetitive) work with robots. Applications of this research can be in education and inclusion of users with special needs. In military and defense, natural and human-like control of mobile robotic systems like drones and ground vehicles may provide leverage over traditional systems. These systems can also be deployed to areas unsafe and hazardous to humans like disaster management and tasks like agriculture, surveillance, search, and rescue. This research may allow operators to have multimodal control of collaborative robots working in manufacturing, agriculture, automobile industries and nuclear power plants.