Publications and Research

Sanyam Garg, Abhinav Tripathi and Edward Cutrell

Accurate Eye Center Localization using Snakuscules

in IEEE Winter Conference on Applications of Computer Vision (WACV '16),

Abstract - Estimating eye centers is an important computer vision problem with several applications. In the past, eye center localization was constrained by the use of special hardware such as infrared cameras. Methods that estimate eye centers based on visible light have also been suggested in the literature, but these methods are inaccurate when used with low resolution images and wide ranges of lighting. We propose a novel method that can be used to detect eye centers efficiently, even with low resolution images provided by a webcam. Our method takes into consideration the circular nature of the iris and its intensity difference as compared to the sclera. It then uses an energy based active contour called Snakuscule for capturing the iris. We test our algorithm’s robustness towards changes in pose, illumination and scale using the BioID database and Yale Face Database B. Our method compares well with existing state-of-the-art techniques in terms of accuracy, is easy to implement and exhibits realtime performance.

Sanyam Garg, Rahul Dubey, Shashikala Nath, Kunda M.M Rao
A Method for Passive Pedestrian Detection at Signalized Crosswalks

in International Conference on Signal Processing and Communications (SPCOM '16), (Under review)

Abstract - Signalized crosswalks can greatly benefit from video system with monocular cameras that detect and counts pedestrians and accordingly vary the waiting / crossing time at intersections. Such systems can also warn drivers about imperceptible pedestrians at night. This paper describes a Histogram of Oriented Gradients and Support Vector Machine architecture that can correctly detect pedestrians at traffic intersections. We evaluate the method’s accuracy by testing it on several manually collected images of pedestrians crossing the road. We further present the detection results on the PETS 2009 dataset and the Town Center video. The algorithm described is able to successfully detect pedestrians in sparse and medium crowd density.

Sanyam Garg, Ramprasaath Selvaraju, Suman Kapur and Kunda M.M Rao

Automated Colorimetric Analysis in Paper Based Sensors [pdf]
in Proc. IEEE International Conference on Image Processing (ICIP '14) 

Abstract - Recent times have seen a wide ranging and large scale use of paper as a potential material in sensors for determining the concentration of an analyte upon appropriate end-point development. Today‘s clinical, food and environmental sectors require low-cost practical analytical devices which are portable and offer on-site real time detection. We present a novel technique for estimating any analyte‘s concentration using a mobile app that analyses the image of the paper subsequent to a chromogenic assay. Making use of snakuscules for capturing the region of interest in the image, followed by basic strategies for removing illumination artifacts using the Von-Kries Coefficient Law, we correlated the luminosity of the colour developed on the paper strip against the analyte‘s concentration. We evaluate our algorithm by determining the glucose concentration levels in blood using commercially available glucometer strips using image processing and comparing them with actual glucose levels as estimated by auto analyzers. The results obtained correlated well with the conventional assay and were almost indistinguishable from the actual values.