Prithvi Raj Singh

Prithvi Raj Singh

Visiting Lecturer @MSU

Actively Seeking opportunities to conduct innovative research as a PhD student.

About Me

Hi 🤚! I am Prithvi Raj Singh 🙂. I am currently a Visiting Lecturer of CS at McNeese State University. For now I am teaching three sections of Begineer and Advanced level JAVA programming as well as Python programming. I've affinity towards topics related with Programming, Computer Vision, TinyML, and Algorithmic Optimization Problems.I graduated with MS in Computer Science from University of Louisiana at Lafayette. I was co-advised by Dr. Raju Gottumukkala and Dr. Anothy Maida.

I finished my Undergraduate from McNeese State University majoring in Computer Science with Minor in Math. I am a avid learner and explorer and I love learning cool new techs. While I am not a fitness freak, I value the importance of regular exercise in mental and physical well-being of person, so I love to do physical activities like lifting weights, running, yoga, and various other fitness exercises

I was born and raised in the southern part of Nepal in well known town 'Birgunj' in working-class family. I spent most of my childhood in my ancestral village about 30 kilometers far from Birgunj. I got my primary school education in my village and moved to town for secondary education. After finishing highschool I travelled to the United States for further education and career opportunities. I love teaching and sharing my knowledge with peers and juniors. I have also been told that I am a great advisor, motivator and co-worker 😉.

Research Interests

As an avid explorer of the vast realm of machine learning (ML) algorithms, my research interests converge on the optimization of cutting-edge technologies in Computer Vision. I am particularly enthralled by the dynamic domains of 3D object tracking and detection, where I contribute to shaping perception systems for robots. During my tenure as a graduate student at UL Lafayette, I delved into the research and implementation of Vision Algorithms, focusing on the intricate task of tracking rapidly moving small objects in 3D-space.

My curiosity extends beyond theoretical pursuits, as I am fascinated by the practical challenges posed by real-world applications. I am passionate about optimizing computer vision algorithms to ensure seamless operations on compact edge devices such as Jetson TX and Raspberry Pi. While contemporary Vision Algorithms excel in detecting and tracking larger, slower-moving objects, they often stumble when it comes to accurately localizing fast-moving, smaller entities like a racquetball. The erratic movements introduce challenges like motion blur and other complexities.

My ambition lies in the relentless pursuit of solutions. I aspire to pioneer advancements in general Robotics perception system as well as improve 3D trajectory analysis of fast-moving tiny objects. To achieve this, I leverage physics-based motion models and delve into the uncharted territories of physics-based deep neural networks. Through my research, I aim to bridge the gap between theoretical innovations and practical implementations, ushering in a new era of precision and efficiency in Computer Vision for real-world scenarios.

While at UL Lafayette, I also worked as a Software Engineering Research Associate for Louisiana Transportation Research Center (LTRC). I helped various Researchers and GeoTech Engineers with implementation of their idea in to usable Software primarily using VB and C# on .NET framework. I am greatly fond of Software Engineering principles and love coding, and learning new tech stack.

To summarize my research interests:

Teaching

Courses I am teaching at McNeese State University for Fall 2024 semester:

Some Projects

Learning with a Limited Sample

As a final term paper for deep learning class offered at UL Lafayette, I conducted experimental analysis on the learning capability of various deep learning based models like CNN, Compact Convolutional Transformers and Vision Transformers. The goal of the research was to understand the efficiency of larger deep neural networks for image classification tasks on various datasets like CIFAR-10, CIFAR-100, MNIST, Fashion-MNIST, and few other small datasets. The result showed that heavier deep neural networks like ViT, CCT are inefficient interms of accuracy and speed, while few layers deep CNN model can give significant accuracy as well as speed. It shows that modern deep neural networks are power and data hungry, and it is not always wise to implement heavier, multilayer neural networks for simple tasks. More details can be at this Github repo.

Natural Language Processing and Image classification with Transformer

In this research, we seek to implement Transformer primarily for the NLP task, focused on Sentiment Analysis, and Language Translation. We performed language translation for English-to-French, English-to-Spanish, Portuguese-to-English. Our general transformer does a decent performance in language translation, and we get acceptable BLEU score, accuracy, and F1 measure for all three language translations. Our sentiment analysis task focused on question answering (Q-A), text classification, and Named Entity Recognition (NER). For the Q-A we can very exact match score of 78%, we got consistent training accuracy of 40% for the NER task. The hugging face transformer API performs wildly better than any present models for NLP tasks. We discuss the recently developed Switch Transformer, MobileViT. Our implementation task for Switch Transformer and MobilViT wasn’t successful as they have huge GPU power. We perform ViT tasks for image classification on CIFAR-10 with a test accuracy of 70% and top-5 accuracy of 97%. Paper and code can be found at this GitHub repository.

Object Detection with Deep Reinforcement Learning

The idea of creating a virtual agent that can follow our instructions and learn from it to do a task we want it to do is fascinating, and the same is the idea for using reinforcement learning in object-tracking problems. We formulate the problem as a sequential decision-making process and the agent is equipped with a conventional object detector that detects the object and the agent is trained to make the detection result better. In this paper, we experiment with using the dynamic method for object detection. Our detector, which uses VGG16 pre-trained on ImageNet as a base, is trained using the PASCAL VOC2012 dataset. The agent takes more actions in transforming the bounding box and the reward accumulated. The AP and recall are best for an IoU threshold of 0.5 or lower. We get AP and recall of 36.68 and 59.0 respectively for the 'cat' class at IoU of 0.5. In our experiment the AP and recall declines sharply for IoU threshold for IoU threshold [0.4, 0.5, 0.6, 0.7, 0.8]. PDF of the paper.

Housing Price Prediction on Kaggle

In this paper, we will be summarizing our work on the Kaggle Housing Prediction competition. We used the D2L book as our reference worked on tuning the hyperparameters and observed the performance of our model. The result of submission on Kaggle for official testing is RMSE. Our work on tuning the hyperparameters didn't necessarily bring any positive change. We have also tried to answer and explain our model, variable standardization, layer count, and regularization. The best performance our system reported was a training error of 0.126395 and, a valid error of 0.146429, but our chart showed that the model was underfitting. We obtained a score of 0.15151 on our official submission on Kaggle ranking 1500th among all participants. Link to PDF.

Error Backpropagation and predicitve coding

This is an implementation of Error Backpropagation (EBP) algorithm which is used in training artificial neural networks. In this research, I conducted comprehensive investigation in to bio-plausibility of the Backpropagation algorithm and deep neural networks. I also explored the innovative applications of predictive coding and assembly calculus in designing a brain emulator that replicates the learning pattern of the human brain. This project was done as term project under the guidance of Dr. Anthony Maida.

Drillobotics: An Automated Drill Rig

In close collaboration with the Petroleum Engineering Department, we worked on design and implementation of an automated drill rig, design which is at the forefront of innovation. We worked on integration of a diverse range of sensors into the rig's architecture which allowed us to get real-time data and perfrom analysis about the precision of the drilling. The prototype design championed the automation in Oil field drilling operations. The project is owned by Petroleum Engineering Department at UL Lafayette.

Omnidirectional Vision System for Continuous Obstacle Avoidance

This project was capstone project for Electrical Engineering Undergraduates whom I pleasure to advise. We worked on building 360 degree vision using PAL360 camera and Microsoft Kinetic Camera. We used ROS2 to program the Turtlebot to move from one location to another and avoid obstacle in the path if any. The Poster of the entire project can be found here.

Thesis

MS Thesis: Real-time Object Detection and Tracking of Fast-moving Small Objects using RGB-D camera and Computer Vision Technicques.

The thesis explores the use of stereo camera for detection and tracking of fast-moving small objects in 3D space. Several Object Detection and Tracking algorithms were experimented. For Object Detection modern YOLO algorithms like YOLOv5, YOLOv7, YOLOv8 as well other mutli and single stage detectors were used. We implemented Kalman Filter based algorithms for the tracking purpose. After several experiments with the state-of-the-art models, the results were unsatisfactory so I worked on augmenting the use of Kinematics to design new tracking system that can enhance the tracking and trajectory forecasting of fast-moving tiny objects. More details can be found at: https://www.proquest.com/docview/3073043126

Undergraduate Capstone Project: Bidirectional autonomous robot for optimal path planning.

Me and my friend 'Blake Drost' worked on building small robot from scratch, that could traverse in forward or backward direction to find the optimal path to reach the target. We implemented breadth first search (BFS) algorithm which the robot used to find the optimal path before making move towards the target. The robot had multiple ultrasonic and small lidar sensors to detect obstacle in the path. If there were obstacle in the optimal path the robot would try to avoid it and if nothing works it will compute new optimal path from the current position and traverse towards that position. More details can be found in our github repository.

Some more Projects related to Exploratory Data Analysis (EDA), Machine Learning can be found on github page. Will continue to add more projects when i get time.

Software Developed

The GeoTechPileCPTSoftware was designed for the Louisiana Dept. of Transportation Development under funding from the Federal Highway Commission. You can download from LTRC's official page.

GeoTechPileCPT

MTS Checker. The software is written in VB.NET and it helps LTRC researchers convert the raw data into usable .txt or excel format. The software is also able to perform data analysis with excel.

Academic and Professional Experience

I worked as a Associate Software Engineer for LTRC for two years maintaining and working on few features, mathematical implementation. We used VB.NET but often we would write code in C# and convert in to VB using Microsoft inbuilt feature in Visual Studio. Some things I learned working there are:

I also worked as RA for Cyber-Physical and Human Systems Lab where I worked mostly on design and implementation of computer vision algorithms for small object tracking in 3D space. I played with a lots of cool tools. Some knowledge i gained are:

References