My interests lie in the broad area of computer vision and pattern recognition. The central question is: What formulations and techniques can extract useful information from images and video. I am passionate about real-world application of computer vision theory. In that sense, I am equal parts a computer scientist and an engineer.
A central theme in my work so far has been the devising of representations for scene quantities of interest that are robust to perturbations that arise from the environment and camera - e.g. viewpoint, illumination, image quality, object variability, etc.
Nokia Research Center, Palo Alto, CA
As of March 2011, I am a senior research scientist at Nokia Research Center. Having worked on virtually every major problem in the visual surveillance domain at Siemens Corporate Research as well as my Ph.D. thesis (see below), I want to explore and learn how the general ideas and principles for algorithm and vision systems design I have developed so far, translate to the domain of mobile computer vision. It is an exciting time for me on several fronts - working on new and challenging problems, working on smartphone platforms, being in Silicon valley, working closely with students from nearby Stanford university, and being part of a new Nokia. I am in charge of developing visual content extraction systems, in collaboration with colleagues from Nokia and NAVTEQ, as well as Stanford University. There will be more to report in due course.Siemens Corporate Research, Princeton, NJ
At Siemens Corporate Research, I was a research project manager, managing projects funded by Siemens Building Technologies and Siemens Mobility. Our team was tasked with building visual surveillance systems that operate under challenging and heavily varying scene conditions while performing acceptably well from the customer point of view. In addition to managing projects, I was a technical contributor, having contributed a suite of algorithms and source code that form the basis for several applications within the SiveillanceTM product family marketed by Siemens Building Technologies (see this link for a press release discussing SiveillanceTM People that I led and contributed to).Design and Validation of a System for People Queue Statistics Estimation, Parameswaran, V., Shet, V., and Ramesh, V.
Illumination Compensation Based Change Detection Using Order Consistency , Parameswaran, V., Singh, M. and Ramesh, V. IEEE Computer Vision and Pattern Recognition Conference, San Francisco, June 2010.
Order Consistent Change Detection via Fast Statistical Significance Testing , M. Singh, Parameswaran, V., and Ramesh, V. IEEE Computer Vision and Pattern Recognition Conference, Anchorage, Alaska, June 2008. (Selected for Oral presentation, acceptance rate 4.0%)
Fast Crowd Segmentation Using Shape Matching, Dong, L., Parameswaran, V., Ramesh, V. and Zoghlami, I., International Conference on Computer Vision, Rio de Janiero, October 2007
Tunable Kernels for Tracking , Parameswaran, V., Zoghlami, I. and Ramesh V.,IEEE Computer Vision and Pattern Recognition Conference, New York, June 2006. (Selected for oral presentation, acceptance rate 4.8%)
My Ph.D. thesis was in the area of visual human motion analysis and more specifically human action recognition and pose estimation. Here is a link to the thesis which includes all of the following contributions:
2D approaches for view-invariant human action recognition
Drawing on results from 2D projective invariance theory, I developed a quasi-view-invariant 2D approach for human action representation and recognition. The straightforward implementation of the key ideas was found to be deficient in many respects and heuristics were designed to overcome these inherent weaknesses. What resulted was an approach for human action representation and recognition that was not only invariant to view-point but also to different subjects, different speeds of execution of a human action, different frame rates and to minor variabilities in the spatiotemporal dynamics of an action. The following two papers discuss the approach.
Quasi-Invariants for Human Action Representation and Recognition, Parameswaran, V., Chellappa, R., International Conference on Pattern Recognition, Quebec, Canada, August 2002.
Using 2D Projective Invariance for Human Action Recognition, Parameswaran, V., Chellappa, R. International Journal of Computer Vision, Volume 66, Number 1, January 2006 )
3D approaches for view-invariant human action recognition
The 2D approach described above has a limitation that the modeled actions need to have a sufficient number of planar poses which is a restrictive assumption if there is a need to model a completely general action. This calls for a 3D approach. It has been established that there are no non-trivial general invariants for 3D to 2D projection. However, that doesn't imply that view-invariant 3D object recognition cannot be accomplished. Model based (or mutual) invariants are one way to accomplish this. Using ideas from mutual invariants based object recognition approaches, I devised a view-invariant action representation and recognition approach which is discussed in the following two papers.
View Invariants for Human Action Recognition, Parameswaran, V., Chellappa, R., IEEE Computer Vision and Pattern Recognition Conference, Madison, Wisconsin, June 2003.
Human Action Recognition Using Mutual Invariants" , Parameswaran, V., Chellappa, R., Computer Vision and Image Understanding Journal, Volume 98, Issue 2, May 2005, Pages 294-324 .
Human Body Pose Estimation for Visual Human Motion Capture
Human body pose estimation is a key initialization step in the use of model based human body tracking algorithms for visual human motion capture. So far, a general solution to this problem has been lacking and this work fills that need. I devised an algorithm based on model based invariants for the recovery of human body pose from a single uncalibrated image with no camera model approximations. The algorithm relies on the isometry assumption (i.e. when appropriately scaled, all humans have the same relative body part lengths) and negligible torso twist. The algorithm appears in CVPR 04:
View Independent Human Body Pose Estimation from a Single Perspective Image, Parameswaran, V., Chellappa, R., IEEE Computer Vision and Pattern Recognition Conference, Washington DC, June 2004
Masters work (CS)
My masters work was in the area of context-based aerial image understanding. The centerpiece of the work was in devising learning approaches for parameter tuning of vehicle and convoy detection algorithms. The work was integrated into the RADIUS (Research and Development for Image Understanding Systems) testbed. RADIUS makes use of context which provides a strong additional cue for image understanding. The key idea is to define a 3D `site-model' which includes a wire-frame model of the major features of the site (i.e. buildings, roads, parking-lots and the like). A subsequently acquired aerial image is registered with the site model, and the site model is projected on to the image, allowing us to delineate regions of interest (e.g. parking-lots). Vehicle and convoy detector algorithms could then be run on those specific regions of interest. The vehicle detector algorithm was based on the Generalized Hough Transform (GHT) and was quite simple. A Canny edge detector is run, followed by votes for candidate vehicle centers based on each edgel's position and orientation. A simple thresholding of votes produces a number of vehicle hypotheses which are verified by template matching (requiring another threshold). For parameter optimization, we used two strategies. Offline Bayesian optimization was done for those parameters which had a smaller impact on the final results. On-line Neyman Pearson based optimization was done for parameters that had significant impact on the results. Specifically, the fact that static vehicles are to be expected at fixed positions in the site is exploited. The parameters are varied to optimize an objective function composed of the non-detection and false-alarm rates. Once the optimal parameters are calclulated, they can be used for the entire image. This makes the algorithms very insensitive to illumination changes in the scene. The following book chapter and paper summarize all of the work:
Site Model Mediated Detection of Movable Object Activities, Parameswaran, V. (with R. Chellappa et al), RADIUS: Image Understanding for Imagery Intelligence, O. Firschein, T. Strat (eds.), Morgan Kaufmann Publishers, 1997.
Performance Analysis and Learning Approaches for Vehicle Detection and Counting Parameswaran, V., Burlina, P., Chellappa, R., International Conference on Acoustics, Speech and Signal Processing, Munich, April 1997.
Masters work (Aerospace Engineering)
In my other life, i.e. before I joined the department of computer science, I was an aerospace engineer and my research area was computational fluid dynamics and the use of numerical techniques for solving partial differential equations as applied to rotary wing aerodynamics and aeroacoustics problems. There were two problems I considered.
Indicial Aerodynamics of Airfoils
The indicial response of an airfoil is the response of its lift to a step change in one of the parameters influencing lift. This is a quantity of fundamental interest for unsteady flows which characterize helicopter rotor blades. Once the indicial responses are known, the response of the blade to an arbitrary temporal variation in the influencing parameters can be found using convolution. While these indicial responses are known for incompressible flow, the underlying equations are too complicated to solve for the compressible case which is the main characteristic for helicopter rotor blades. For small and large times, exact analytical expressions can be derived for indicial responses. However, for the more important intermediate times, no solution existed before my work. My work was the first that formulated the problem in a way that could be solved using computational fluid dynamics. The results agreed very well with known solutions for small and large times. The approach was extended for the 3D case using the same fundamental idea and eventually led to a Ph.D. and Masters theses for subsequent students in the lab. The work is described in the following paper:
Indicial Responses in Compressible Flow, Direct Calculations, Parameswaran, V., Baeder, J.D., Journal of Aircraft, Vol. 34, No. 1, Jan-Feb 1997.
Helicopter Blade Vortex Interaction Noise
Vortices (masses of spinning air) are an inescapable consequence of the production of lift on an aircraft, be it a fixed-wing airplane or a helicopter. The vortices that trail helicopter rotor blades, linger and interact with the blades as they rotate, producing impulse loadings on the blades which in turn produce noise. This is called Blade Vortex Interaction (BVI) noise. There has been interest in producing quieter helicopters for various reasons. One way that I explored was the pitching of helicopter blades (using piezoelectric actuators) in order to counteract the impulse loading produced by a passing vortex. The key idea was to consider the change in lift produced by a moving line vortex and assume that it is the result of an equivalent change in angle of attack of the blade. This formulation results in a simple ordinary differential equation that the angle of attack satisfies which can be solved using straightforward finite difference equations. Once the angle of attack history (alpha(t)) is calculated, the blade is pitched to negate it (-alpha(t)). Computational fluid dynamic calculations showed a large reduction in noise for such a pitching blade. The approach is described in the following AIAA conference paper.
Reduction of Blade Vortex Interactions using Prescribed Pitching, Parameswaran, V., Baeder, J.D., 13th AIAA (American Institite of Aeronautics and Astronautics) Conference, San Diego, June 1995.