Robotics and Machine Intelligence Laboratory

Mentor: Prof. Rajesh Rohilla

Real-time object recognition system for autonomous navigation. Used the YOLO architecture and ZED stereo camera on Nvidia’s Jetson TX1 module to develop a system that provides real-time information about the identity of objects and their distance from the car using audio.

Implementing YOLO in Pytorch from scratch: Code

Road Traffic Counting (Digital Image Processing Course Project)

Prachi Garg, Ashi Gupta, Prof. Dinesh Chutani

In this project, we implement a motion detection and tracking algorithm to monitor road traffic passing through the road as captured by a video camera. We use OpenCV and Python to implement motion detection and tracking using background selection algorithms. Project Page

Machine Learning projects:

• Naive-Bayes for newsgroup classification in the 20 Newsgroups dataset. I used to compare the Scikit Learn implementations with the original algorithms. Code

• I love understanding the mathematics behind machine learning models and algorithms. I used to compare popular algorithms on the Scikit-Learn datasets and infer what kind of algorithms work best for which type of datasets and why. For example, there are implementations of Voronoi and PAM K-medoids algorithms, along with the Sklearn inbuilt ones. Code for this ML repository from when I first started learning: Code

• For my first basic project to understand CNNs and hyperparameter tuning, I collected a dataset of dogs from my colony. Then I implemented various CNNs to train a Dog classifier. I also used YOLO to train a Dog detector on my dataset. Classifier Code / Detector Beginner Tutorial

Algorithm to find all the possible spanning trees of a given graph in increasing order of cost.

Solved using Prim's method: Code

IRL Technoxian Quadcopter Challenge, World Robotics Championship, by International Robotics Council (IRC), 2017

We built an indigeneous Light-Weight Sandwich Composite Multirotor System. As the Avionics Technician on the project, I was responsible for designing, integrating and testing the quadcopter's propulsion system and tuning the Naze flight controller. Report

AUVSI Student Unmanned Air Systems Competition, Maryland, USA, 2017

UAS Lazarus

Journal Paper (Team UAS-DTU)

Developed a customized Ardupilot firmware algorithm to capture the off-axis target using 2-axis gimbal stabilization for Lazarus. Did extensive flight testing of this feature for the AUVSI-SUAS competition.