- I worked on this project with Dr. Xiang Liu as part of my job at Center for Advance Infrastructure and Transportation (CAIT).
- For this project I developed a computer vision program to analyze the traffic from CCTV footage for Center for Advanced Infrastructure and Transportation.
- The program would take a video file consiting CCTV footage at a railway crossing and computed count, relative speed, direction and time of crossing for both pedestrians and vehicles.
- The program achieved accuracy of 94% and computation time 60% less than actual video time, making it suitable for real time application. * The project was oriented on creating computationaly light weight video processing solution which can potentially be used on light weight systems (like raspberry pi) in real time.
- The program used background subtraction and Kalman filter for tracking and counting.
Github : As this project is confidential, the Git repository is private. I have asked for permission to share some details of the project.
Understanding Crowd Behavior using Unsupervised Deep Neural Networks
- For this project we deployed two individual models for simulating crowd movement behavior in public places, such as subway stations, in a team of four.
- The proposed approaches include non-linear PCA based networks belonging to the autoencoders family more specifically Variational Autoencoders, as well as deep generative models trained under an adversarial setting more specifically WGAN.
- I was soleley responsible for representation and feature extraction, where I created a novel representation which used probabilistic heat map generated with overlapping Gaussian kernels.
Project Report :
Proposing a Novel Method for Fake News Classification
- In this study after exhaustive study of current methods for fake news detection, we came up with ensemble model consisting two classifiers.
- First was a model with an information retrieval module and a feed forward neural network to integrate the knowdlegebase for fake news detection. Second was style based classification model utilising word embeddings and Bi-LSTM.
Project Report :