Research
Identifying Temperate Exoplanet Candidates in the TESS Continuous Viewing Data
Research Internship at Academia Sinica Institute of Astronomy and Astrophysics, Taiwan. Supervised by Dr. Alex Teachey
The TESS CVZ is the overlap region between the various sectors of the TESS all-sky coverage, which provides us a time baseline of around 351 years, enabling us to search for longer period exoplanets in the previously (relatively) undiscovered CVZ data. We created a pipeline to search for exoplanets using a detrending mechanism (using Median filter and Tukey's Biweight filter), an unsupervised Machine Learning based outlier rejection mechanism, and a Box Least Squared transit search module. We also used a grid-based approach to size our kernels for the transit search which were informed a priori by stellar parameters.
The pipeline was able to recover 18/20 known Kepler planets and 17/20 known TESS planets in the test dataset. We are continuing this research on finding and following up for CVZ planet candidates
Paper in preparation
Project Proposal: Exoplanet Transit Classification using Recurrent Neural Networks
Supervised by: Howard Isaacson, Research Scientist at UC Berkeley, and mentored by Joseph M.A. Murphy, Grad student at UCSC
Submitted to Intro to Astro 2020 course
Exoplanet_TransitRNN_Research_Proposal.pdfDriver Assistance Systems - Vehicle to Vehicle and Vehicle to Pedestrian Collision Avoidance
Research Internship 1: Marconi Society, under the supervision of Dr. Aakanksha Chowdhery, Google Brain, Tensorflow (then Postdoc at Princeton) and Prof. Brejesh Lall, Electrical and Electronics Dept., IIT Delhi and mentorship by Prerana Mukherjee, Grad Student at IIT Delhi and Abhishek Gagneja, Assistant Professor, Bharati Vidyapeeth's College Of Engineering
Research Internship 2: IIT Delhi, under the supervision of Prof. Brejesh Lall, Electrical and Electronics Dept., IIT Delhi, and under mentorship from Abhishek Gagneja, Assistant Professor, Bharati Vidyapeeth's College Of Engineering
Application based solution for collaborative driver assistance system
drizy_poster_mobicom17.pdfMobiCom_17.pdfDataset
Collected and annotated a dataset of over 50,000 frames with 7 different classes of road obstacles (including cars, pedestrians) along with occlusion information. The dataset was the only indigenous dataset with the above information available and was leveraged for benchmarking various detection and tracking algorithms
3_Annotations.mp4Multi-Object Tracking
In Multi-Object Tracking, each of these classes' objects in the frame were treated as separate objects with given IDs, and algorithms were used to track their trajectory through the frames for each of these objects. The main algorithms in use were SORT (Simple Online Realtime Tracking), and the Neural Network implementation of SORT, called Deep SORT, which learns features to capture the trajectory of these objects. These algorithms were implemented on the indigenous dataset to optimize the performance and benchmark the dataset for Multi-Object Tracking.
Fig: A frame from the SORT Tracking for multiple objects in the frame
Trajectory Prediction: Survey
In Trajectory Prediction, we focus specifically on pedestrians and try to predict the future trajectory points for them based on their previous trajectory. We surveyed the field extensively to implement algorithms such as Social LSTM, SS-LSTM, and Social GANS, among others. We also researched novel ways to implement trajectory prediction for the dynamic dashboard cameras which had no published and open algorithms at that time,
Visible Light Communication
A visible light communication (also known as LiFi) prototype was created as part of the idea to enable short-term underwater communication using predetermined message signals. The paper presents our results on the time-delays and feasibility of the prototype. Slides can be seen here
Aquacom.pdf