|
Multiobjective Robust Control of Nonlinear Systems (slide show (pdf)).Future control systems will be required to deliver near optimal performance operating under a wide range of conditions. Thus, these controllers will necessarily have to deal with the nonlinear features of the plant. In this project we are developing a simple synthesis framework, based upon the combination of Receding Horizon and Control Lyapunov Functions techniques that explicitly takes into account multiple performance specifications and model uncertainty . This framework has been validated using simplified models of thrust vectored aircrafts, where the resulting controllers achieved virtually optimal performance, significantly outperforming controllers designed using classical techniques. |
|
Robust Identification and Model (In)Validation of LPV SystemsDuring the past few years efficient tools have been developed to robustly stabilize Linear Parameter Varying systems. However, a key issue that needs to be addressed in order to apply these techniques to practical problems is the development of techniques that generate and (in)validate suitable models. In this project we are developing a control--oriented identification and model (in)validation framework for LPV systems that takes into account both the dependence of part of the model on time--varying parameters as well as the possible existence of a non--parametric component. Our main result shows that these problems can be recast as a Linear Matrix Inequality feasibility problem and efficiently solved. Moreover, the overall computational complexity is similar to that of obtaining and/or (in)validating LTI models of comparable size. These results have been validated in several practical examples arising in the context of active vision . |
|
System Theoretic Tools for Physics. (slide show (pdf))This project seeks to apply concepts and mathematical techniques, arising mainly from systems and control theory, to foundational questions about physical systems. Even though the traditional wisdom usually connects Systems Theory and Physics through the many successful engineering applications where both play an important role, there are enough convincing arguments to argue that the links can be made much stronger, and more direct. An example of the potential of this approach is its ability to offer simple explanations to some persistent ``mysteries'', such as how microscopically time--reversible dynamics lead to apparently irreversible macroscopic behavior. Simple model reduction arguments show that this "irreversibility" originates as an artifact of partial observations over finite time intervals. |
|
Operator Approach to Robust Multiframe Tracking. (slide show (pdf)).Conventional algorithms for tracking a target in a sequence of images can fail due to a combination of clutter and occlusion . In this project we are developing a new operator--theoretic framework to robust tracking. By exploiting both spatio--temporal and a--priori information , the resulting algorithms have proved to be substantially more robust than the current state of the art. (example1, . example2, . example3) . Further robustness improvements can be achieved by tracking suitable sets of parts (example4: tracking using parts) and by modeling how the target appearance and size changes over time. |
|
Robust Active Tracking. (slide show (pdf))Recent hardware advances have rendered active vision a viable option for a very diverse spectrum of applications ranging from MEMS manufacture to surveillance and assisting individuals with disabilities. However, there are relatively few instances where these techniques have been successfully applied in uncontrolled environments. This can be traced, to a large extent, to the fragility of active vision systems designed using classical methods. (example1: tracking a car using a PID) . (example2: tracking a person using a PID) . In this project we are using a combination of robust identification and control synthesis techniques to synthesize robust active vision systems that can accommodate substantial uncertainty, stemming for instance from uncertain time delays, unmodelled dynamics and changing optical parameters (example1: robustly tracking a car) . (example2: robustly tracking a person) . |
|
Robust Tracking using Parts. (slide show (pdf)).Tracking an object in a sequence of images can fail due to partial occlusion or clutter. Robustness can be increased by tracking a set of parts , provided that a suitable set can be identified. In this project we propose a novel segmentation, specifically designed to improve robustness against occlusion in the context of tracking. The proposed segmentation is obtained by driving a snake segmentation to enclose regions with good tracking properties, such as texture and shape. We have shown that tracking these regions outperforms both tracking parts obtained through traditional segmentation algorithms and tracking the object as a whole . |
|
Structure from Motion and Dynamics (poster (pdf))Traditional structure from motion algorithms are based on the factorization of a matrix whose rows contain the position in the image of the features tracked over time. While these approaches lead to elegant, computationally efficient solutions, they remain fragile to noise, missing correspondences and large inter-frame motions. In this project we use Caratheodory- Fejer interpolation (CF), recently developed in the control field, to estimate the motion dynamics based on a set of previous frames. The estimated dynamics are then recursively used to robustly estimate object structure and motion and to predict the object appearance in future frames. Results are tested for 3d reconstructions of rigid bodies in real and synthetic image sequences. | From
|
Pedestrian Detection and Gait Recognition. (slide show (pdf))Biometrics is the science of automatically identifying individuals based on their physiological or behavior characteristics. One such behavior characteristic that allows to identify a person at a distance is the way the person walks . In this project we are using algebraic and operator-theoretic tools to learn and recognize the gaits of a set of individuals. The proposed approach uses small numbers of frames to extract samples of the ``shape of the motion'' from walking sequences and compresses them into a few representative samples. Individual and activity recognition is achieved by finding the exemplar closest, in a suitable metric, to the data being observed. The small number of frames maximizes the training data and at the same time increases robustness to occlusion. |
 |
Systems Theoretic Tools for Robust Texture Inpainting, Synthesis and Classification slide show (pdf)The objective of this research is to develop a comprehensive approach to image texture that addresses several key subproblems – modelling, inpainting, classification and segmentation – in a common framework. Its conceptual backbone is a systems theoretic viewpoint that emphasizes both robustness and computational complexity issues and practicality of the results. The main idea is torecast these problems as the search for an operator satisfying suitable interpolation and structural constraints. |
|
Obstacle Detection for High Speed Civil Transportation (HSCT)(This is joint work with R. Kasturi and L. Coraor )NASA in collaboration with industry is developing a supersonic passenger airplane, called the High Speed Civil Transport (HSCT). Due to aerodynamic considerations, the nose of the new airplane will not allow pilots to look at the runway during take-offs and landings, unless it is drooped during these maneuvers. An alternative solution that is being examined is to replace the forward cockpit windows with synthetic displays. The imagery in these displays would be obtained from video cameras mounted outside the aircraft. As part of this project, we are developing computer vision algorithms to process video captured from an aircraft to alert pilots of possible obstacles on runways , on a crossing path in front of the aircraft, and on a collision course. |
|
Object Recognition in the Presence of Clutter(This is joint work with T. Kanungo)Recognition of objects among clutter requires a good representation capable of handling occlusion and segmentation problems for a large number of free-form objects. In this project we use a representation based on the appearance of parts and relations which are obtained from segmentations of training images and are compressed using Karhunen-Loeve (KL). Since this representation is learned from segmented images, it is capable of representing free-form objects and handling segmentation problems similar to the ones encountered during training. Furthermore, since it is based on local parts rather than on global properties, it is robust to partial occlusion . |
|
Omnidirectional VisionApplications such as autonomous navigation and visual surveillance are restricted by the narrow field of view of traditional sensors. Catadioptric sensors combine mirrors and CCD cameras to provide 360 degrees images. Unfortunately, this can only be achieved using mirrors with curved surfaces that result in non-uniform image resolution. In this project we are studying the problem of aligning multiple omnidirectional images to improve image resolution, compute absolute depth of the scene and track moving objects in front of the sensor. |
