Projects

IMG_7655.MOV

IMG_7738.MOV

When do humans in VR take over Autonomous car in off-road conditions? An IRB.

This study investigates human reactions to autonomous vehicles in off-road driving conditions. Participants will engage with a virtual environment created using the CARLA driving simulator, presented through a VR headset. The study aims to observe and analyze scenarios where individuals decide to overtake an autonomous vehicle and resume manual control. Specific conditions that influence these decisions, such as terrain complexity, vehicle behavior, and environmental factors, will be examined. Data from the simulation will contribute to a human behavior survey to inform future autonomous driving system designs.

AI’s take on history through sedimentology of paleo periods of Earth

This project leverages artificial intelligence to analyze and interpret the sedimentological data of paleo periods of Earth, offering a novel approach to understanding historical geological events. By collecting over 40,000 research papers related to paleo-sedimentology, the project systematically extracts critical information such as event location, methodology, and key findings. Utilizing advanced tools like large language models (LLMs), PyTorch, web scraping, BeautifulSoup, and Retrieval-Augmented Generation (RAG), the system processes and synthesizes vast amounts of research data to uncover patterns and insights about Earth’s ancient geological history. The integration of AI-driven techniques allows for a more efficient and scalable analysis of complex sedimentological data, facilitating deeper insights into the Earth’s paleoenvironment and supporting future geological research.

Sequential activities of brain. Mapping brain activities using Machine Learning

This project explores the sequential activities of the human brain by developing a machine learning model that maps electroencephalography (EEG) and magnetoencephalography (MEG) brain activity to specific actions. The model captures the temporal dynamics of brain signals to predict and identify the corresponding actions being performed. To enhance the accuracy and applicability of the model, real human data was collected, ensuring the project remains grounded in real-world conditions. By leveraging advanced machine learning techniques, the project aims to advance our understanding of brain activity patterns and their correlation with cognitive and motor tasks, potentially contributing to fields like brain-computer interfaces (BCI), neuroscience, and cognitive rehabilitation.

Which actions did the gymnast perform? Predicting Sub-actions for Action Quality Assessment in Gymnastics

This project introduces a novel approach to predicting sub-actions performed by gymnasts in video footage, aimed at improving action quality assessment in gymnastics. By combining computer vision and natural language processing (NLP), we developed a Vision+NLP-based model capable of identifying and classifying sub-actions in dynamic gymnastic performances. The model analyzes video data to detect intricate movements, while NLP techniques enhance the contextual understanding of these actions. A key contribution of this work is the creation of a new dataset specifically designed for the task, which facilitates the training and evaluation of predictive models in the context of sports performance. Implemented using Python and PyTorch, the project leverages state-of-the-art tools in image/video processing, deep learning, and database management systems. The research, conducted in part at IIIT Hyderabad, is currently under submission for publication, contributing to the growing intersection of vision and NLP for action recognition in sports.

DeepSAR: Vessel Detection in SAR Imagery With Noisy Labels

Unlike traditional optoelectronic satellite imaging, Synthetic Aperture Radar (SAR) allows remote sensing applications to operate under all weather conditions. This makes it uniquely valuable for detecting ships/vessels involved in illegal, unreported, and unregulated (IUU) fishing. While recent work has shown significant improvement in this domain, detecting small objects using noisy point annotations remains an unexplored area. In order to meet the unique challenges of this problem, we propose a progressive training methodology that utilizes two different spatial sampling strategies. Firstly, we use stochastic sampling of background points to reduce the impact of class imbalance and missing labels, and secondly, during the refinement stage, we use hard negative sampling to improve the model. Experimental results on the challenging xView3 dataset show that our method outperforms conventional small object localization methods in a large, noisy dataset of SAR images.

Seizure Prediction using EEG Data and Deep Transformer Models and CNN+LSTM

Bachelors of Science (Undergraduate) Thesis. View Project.

SNAKE-BOT :

A robotic structure resembling a snake that can move according to the given environment. This can be used for excavation, security purposes, and detection of drug trafficking under the land etc.

This project won the Rajasthan Hackathon and was funded with ~ $22K / Rs. 15 Lakh. For Twitter click here.

HANDY :

The project aims to lower the barrier in communication with the people who are deaf and mute. Traditionally, flex sensors have been used to detect the hand movements. But it is inaccurate and gives the same value no matter where the bent is. To solve this problem, we placed potentiometers at every joint of our hand so that we can detect exact hand movement and convert sign language in speech(or text). The person has to wear the glove and he can communicate using sign language. His phone needs to have the application that will generate corresponding speech that the other person can listen.

This project won the BITS Mesra Hackathon (HACK-A-BIT). Click here. Click here.

NUTTY SQUIRREL :

In this theme, a robot and a lift mechanism is designed that depicts a Squirrel that sorts, carries and places nuts at different places on land and in trees. The challenges in this theme includes V-REP Simulation, Path Planning, building a Robot and Lift Mechanism, Sensor Interfacing, designing a mechanism to pick and place the Nuts. A configuration image will be given to teams. The arena consists of thermocol blocks of different colours that are placed randomly. The robot traverses the arena to pick up a block and place it at a designated section S1 and S2. The teams needs to make a lift mechanism to carry the bot at a certain height to place blocks at the designated section S3.

This project was in top 20 of IIT-B National Robotics Competition e-Yantra. Click here.

OPTICAL FLOW :

A feature tracking algorithm using Lucas - Kanade method. By combining information from several nearby pixels, the Lucas–Kanade method can often resolve the inherent ambiguity of the optical flow equation. It is also less sensitive to image noise than point-wise methods. On the other hand, since it is a purely local method, it cannot provide flow information in the interior of uniform regions of the image.

This was one of my task during my internship for feature tagging and its location.

Object Detection :

Tried to implement Mask RCNN algorithm on Indian conditions.

ATTENDANCE SYSTEM :

Designed a prototype system for college to record the attendance of staff and students at same time using Bio-metrics.

Google Sites

Report abuse