Nishu Choudhary

I am a Ph.D. candidate in Traffic Operations Research at Georgia Tech advised by Dr. Michael P. Hunter and Dr. Angshuman Guin .

My research uses Machine Learning (ML) techniques to predict traffic congestion enabling Traffic Management Centers (TMCs) to implement real-time traffic mitigation strategies such as ramp metering and speed harmonization, leading to more energy-efficient transportation system.

Email  /  CV  /  Bio  /  Google Scholar  /  Twitter  /  Github

I'm interested in exploring the potential of Machine Learning in providing data-driven systems level solutions for supply chain systems.

Machine Learning approaches are often considered to be "magical" black box algorithms that have the ability to learn any type of pattern. While this may be true to some extent, the paper provides evidence that in order for ML algorithms to learn precursors of demand-related traffic congestion, they require information from the most downstream (active) bottleneck. In other words, to predict congestion, features from the source of congestion, which is the active bottleneck, are necessary.

As a part of implementing an end-to-end solution for Proactive Traffic management Traffic Congestion index is proposed. The index capitalizes on information provided by the density of indicators of traffic instability to indicate probability of congestion.

This project is a joint effort from North Carolina State University, Georgia Tech, and Florida International Univeristy funded by Southeastern Transportation Research, Innovation, Development, and Education Center (STRIDE). The project aimed to help transportation agencies use “big data” to mitigate congestion and improve system performance for both freeways and arterials.

The comparison revealed challenges related to noisy data and the failure of spot-speed measurements to adequately capture heterogeneity in congested traffic, which rendered delay estimation methodologies impractical for field use. To address this challenge, a regression model was developed using data from a non-exhaustive set of incident scenarios simulated using Vissim®, to help obtain rapid estimates of delays for incidents with varying characteristics occurring under varying base conditions.

This project evaluates the feasibility of the use and potential benefits of a video-based automatic incident detection (AID) technology relative to existing detection via the Georgia 511 (NaviGAtor) system. The project also investigates the potential of crowdsourced smartphone app-based incident detection and notification, in reducing the time to detection.

Blocking of the box can lead to increased travel times and, in extreme cases, to gridlock. This research investigated the effectiveness of “Don’t Block the Box” (DBTB) treatments in minimizing driver’s blocking an intersection by 1) Field study conducting a “before-after” DBTB treatment comparison study at six intersections and 2) a simulation study.

This paper presents a python-based software application “GT-MVP” designed to provide a user-friendly interface to collect complex video-based traffic data. GT-MVP’s graphical user interface allows users to play multiple videos and operate them synchronously using common controls, to review the extracted data in real-time, and to correct errors easily.

1. From scratch (only NumPy) implementation of RNN


2. From scratch implementation of Backpropagation


3. LeCun et al. (1989) (PyTorch)


4. LeNet 5 (PyTorch)


5. Bengio et al. 2003 (NLTK, PyTorch)

Page Template credit: Dr. Jon Barron.