Tejoram Vivekanandan

I am a Research Assistant at the Computational Imaging Lab of Indian Institute of Technology, Madras. My advisor is Dr. Kaushik Mitra.

Prior to this, I have had the privilege of working with Dr. Glenn Orton as a Research Fellow at the NASA Jet Propulsion Laboratory under the Juno mission.

I previously worked as a Research Intern with Ms. Thara Nair at Indian Space Research Organisation, Hyderabad. I completed my undergrad in Electronics and Communication Engineering at the Coimbatore Institute of Technology.

Email  /  CV  /  Github  /  LinkedIn

I'm interested in computer vision, machine learning, and image processing. Much of my research is about inferring the physical world (shape, motion, depth, color, light, etc) from images and videos. I also like to focus on interdisciplinary research, applying vision-based methods to innovatiely tackle problems in different fields. More recently, I have been working on perception in self-driving vehicles.

To determine the correlation between the color change, the altitude and ultraviolet exposure of the storm’s particles, near-infrared images of Jupiter with Oval BA present were examined. After calibration and preprocessing the images during red, white and the transition phase were compiled for different wavelengths. The reflectivity of the vortex at each wavelength was adjusted subject to angles of emission and incident sunlight using the Minnaert function. Results confirmed a change in altitude of particles between the red and white epochs.

Shadow restoration approach for high resolution satellite images was adopted. This approach detected the shadow area and segmented the image into regions relevant to the types of land surface. Thereafter, shadow restoration was applied region-wise in relation to the degree of correspondence between shadow and neighboring non-shadow regions. The results proved that the shadow regions processed had a better appearance and were highly compatible with surrounding non-shadow regions. Thus, the final accuracy was more than that of the conventional approaches.

A prototype of an intelligent self-driving vehicle was developed on a Raspberry Pi with a variety of machine learning algorithms. It was able to predict the direction and control the vehicle based on the predictions. The prototype was trained to recognize traffic signs and to navigate without collision. For this purpose, the images of a track collected from a Pi camera were used to train different models of neural networks and the performance of each model was tested. Haar cascade classifier based stop sign detection signals were used to stop the vehicle. 95% decision accuracy was attained using softmax activation function with 256 hidden layer nodes.

In this project, handwritten numerics from a live video input was recognized and matched with telephone directory to find the name associated with the detected telephone number. Image background was removed through edge detection, localization and perspective transform. After Thresholding, ROI bounding boxes were computed. Finally digits were recognized by deep neural network through classification.