Research Projects

Some of the projects that I have done in recent times are briefly described here along with links to their blog posts, reports and github repositories.

Yoga Activity Estimation for Real Time Video
(December 2017 - August 2018)

python posture estimation keras LSTM CNN

• The system takes estimated poses from Openpose and uses them to train a deep learning model for yoga posture estimation
• Achieved 99.32% accuracy on initial data
• Realtime implementation achieved 98.92% accuracy
• Submitted for publication
• Demo

Deep Convolutional Neural Networks for Traffic Sign Recognition & Detection
(June 2017 - July 2017)

keras theano deep learning gtsrb gtsdb vgg vgg16

• A custom 2D Convolutional Deep Learning Model inspired by VGG architecture with additional batch normalization layers was designed
• Majority voting was applied for ensembling for increased robustness and accuracy
• German Traffic Sign Recognition Benchmark [GTSRB] dataset was used for comparison with state of the art techniques
• Achieved 99.38% classification accuracy on the 43 categories of traffic signs
• Ranked 7th internationally on the GTSRB benchmark, beating any human performance based approach and achieving results comparable to academic state of the art
• My blog post explaining the model, ensembling and results along with some suggested approaches for further research
• Source code [Report]

Neural Networks for Stock Price Prediction
(August 2017 - December 2017)

python keras multimodal multitask LSTM cnn deep learning financial forecasting stocks stock market

• Designed a Multimodal and Multitask Deep Learning Model to predict stock price movement and volatility
• Price History and News headlines were taken as two modalities and used as inputs to two separate LSTMs
• Outputs of both LSTMs were merged and two independent branches of dense layers were used to predict price movement direction and volatility
• Achieved a prediction accuracy of 74.93% for predicting price movement direction and MSE of 0.0017 for normalized volatility prediction
• Read my blog post for insights on this work
• Source code [Report]

Semantic Similarity Analyzer and Predictive Typing System
(October 2017- November 2017)

python keras Siamese LSTM Manhattan LSTM MaLSTM Semantic

• Used Manhattan LSTM to predict semantic similarity of two query phrases
• Google word2vec was used to generate embeddings of query phrases
• Achieved an accuracy of 80.35% on Quora Question Pairs Dataset
• Semantic similarity between current sentence and sentences in the corpus was used for predictive typing
• It outperformed state of the art statistical techniques
• Read my blog post for design, implementation and results of the Semantic Similarity Analyzer.
• Source code [Report]

Compiler Construction
(January 2017 - December 2017)

compiler programming languages C Data Structures

• A compiler for a language supportibg arithmetic, string and matrix operations was developed in C using hash tables, trees and stack optimally.
• Completed as a part of Compiler Construction course from Jan - May 2018 at BITS Pilani
• Repo

ATU: Adaptive Template Update for Constructive Fingerprint Identification
(January 2017 - December 2017)

python menpo numpy scipy thin plate spline tps minutiae biometrics fingerprint

• Minutiae points were mosaicked using Thin Plate Spline transformation
• It mimiced natural deformation during registration of fingerprints and counter acted them to achieve consitency between different fingerprints of the same user
• High confidence test fingerprints were used to update the database resulting in knowledge augmentation and efficiency
• Achieved 100% CRR and 20% EER for fingerprints of 100 users of FVC 2004 database
• Source code [Paper]

Smart AC System

• Designed Smart AC System using 8086 microprocessor, RAM, ROM, PPI, Timer, LCD, address latches & buffers.
• Project was simulated successfully using Proteus
• Source Code Report

Logic Programming using Prolog

• Logic Programming using SWI Prolog to implement Symbolic Algebra Operations
• Designed Programming logic to implement BITS Academic Regulations in Prolog
• Project Repo with Prolog source code and documentation