Clint K. Morris

Machine Learning Graduate Assistantship

Smart Mobility and Infrastructure Lab

RESEARCH OVERVIEW:

Clint Morris demonstrated exceptional proficiency by first authoring five academic journal publications and co-authoring two additional studies during his pursuit of a Master's degree. His research was conducted at the University of Georgia's Smart Mobility and Infrastructure Laboratory, a cutting-edge facility dedicated to addressing contemporary mobility and infrastructure challenges. The laboratory emphasizes engineering, data science, machine learning, deep learning, computer vision, edge/cloud computing, sensor networks, and remote sensing to develop innovative solutions in the field.

First Authored Publications:

• Accident Analysis and Prevention: Morris C. and Yang J. (2021). Effectiveness of Resampling Methods in Coping with Imbalanced Crash Data: Crash Type Analysis and Predictive Modeling AAAP-D-20-00487.
• Machine Learning with Applications: Morris C. and Yang J. (2021). A Machine Learning Model Pipeline for Detecting Wet Pavement Condition from Live Scenes of Traffic Cameras. MLWA-D-20-00413.
• MDPI Infrastructures Journal: Morris C. and Yang J. Understanding Multi-Vehicle Collision Patterns on Freeways—512 A Machine Learning Approach. Infrastructures 2020, 5(8), 62; https://doi.org/10.3390/infrastructures5080062

Pending Publications:

• Transportation Research Board: Ma S., Yang J. Chorzepa M., Morris C., Kim S., and. Durham S. (2021). A Composite Deep Learning Architecture for Vehicle Classification – Combining Vision Transformers with Wheel Detectors.
• Transportation Research Congress: Morris C. and Yang J. (2021). A multi-source hierarchical deep learning-based self-supervised architecture for quality control of traffic data.

Link to Laboratory Webpage Close Window

Data Science Internship

Charging and Data Modeling Team

Data Science - Algorithms:
Worked within the Charging and Data Modeling Team at Tesla, Inc. Developed Causal Inference Machine Learning Models to Optimize SuperCharger Placement centred on maximization of ROI.
Created network based forecasters to asses and predicted impact of neighboring SuperCharger sites.

DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Causal Inference Machine Learning Models
 • Network Modeling
 • Advanced Forecasting
 • Statistical Modeling

Close Window

Machine Learning

Algorithms Engineer

WORK OVERVIEW:
 
Research and Algorithms Engineer at the Georgia Tech Research Institute (GTRI), specializing in the Electro-Optical Systems Laboratory within the Electronic Warfare Modeling and Analysis Division. Expertise in developing and implementing innovative machine learning techniques and conducting comprehensive data analysis to enhance algorithm performance. Adept at executing rigorous algorithm tests, designing and refining algorithm architectures to align with and advance military defense objectives.

HIGHLIGHTS:
 • Obtained Security Clearance: Secret
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Automated Machine Learning
 • Advanced Transformer Architectures
 • LiDAR
 • Object Tracking
 • Object Detection
 • Genetic Programming

Link to Laboratory Webpage Close Window

Clint's Resume

Link to Resume Close Window

Machine Learning

National Fire Control Symposium (NFCS)

WORK OVERVIEW:
Clint developed configurable software for optimizing detection and classification in fused Automated Target Recognition (ATR) for the National Fire Control Symposium (NFCS). The NFCS, a collaborative forum, addresses kill chain requirements, translates needs into technology specifications, identifies challenges, discusses priorities, maximizes return on investment, and reports progress on current investments.
 
HIGHLIGHTS:
 • Clint presented and published his research on cutting-edge detection technology, showcasing his expertise and contributions to the field.
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Novel Transformer Architecture
 • Deep Learning Frameworks (Tensorflow and Pytorch)
 • Innovative loss functions to manage significant intra-class and inter-class variations effectively
 • Effectively managing noisy labels with efficient techniques

Link to Webpage Close Window

Machine Learning

Preprocessing Steps for Multi-Modal Sensor Algorithmic Fusion for Sensor Week ’18 Data Collection to Improve Machine Learning Model Performance

WORK OVERVIEW:
(U) Limited multi-modal sensor fusion datasets, particularly those relevant to the Department of Defense, have varying label quality, hindering the effective use of machine learning applications. The creation of labeled datasets often requires extensive human involvement, which can be inefficient in terms of time and cost. To address this challenge, our research aims to develop methods to improve label quality with minimal human involvement, focusing on the Sensor Week '18 dataset.

(U) We examine three categories of truth label enhancement: optimal dataset distribution, pixel-space correction, and geo-space correction. To prevent model overfitting and optimize dataset distribution, our research explores methods to minimize dataset similarity. Proposed pixel-space correction methods include unsupervised techniques such as k-means clustering, fuzzy C-means clustering, and filter bank clustering with rule-based shape analysis, as well as a semi-supervised middle-out chain label correction approach utilizing low-level object tracking.

(U) Geo-space errors within the dataset are rectified using a multi-stage homographic transformation method, which incorporates unsupervised matching of ground control points to existing satellite imagery. Our research demonstrates significant improvements in detection and classification accuracy across multiple sensing modalities by training machine learning models with corrected truth data, and also explores the degree of human supervision and annotation required for each method.
 
HIGHLIGHTS:
 • Military Sensing Symposium (MSS) and Battlespace Acoustic, Seismic, Magnetic, and Electric-Field Sensing and Signatures Committee (BAMS).
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Deep Learning Frameworks: PyTorch and Tensorflow
 • Novel Local Patterns Feature Extraction Algorithms: SIFT, SURF (Speeded-Up Robust Features), ORB (Oriented FAST and Rotated BRIEF) and Boosted Efficient Binary Local Image Descriptor (BEBLID).

Link to Webpage Close Window

Machine Learning

Understanding Multi-vehicle Collision Patterns on Freeways – A Machine Learning Approach

RESEARCH OVERVIEW:
In this research project, various forms of collision and weather data were generate and coupled for machine learning model inffernece.
 
HIGHLIGHTS:
 • First Author
 • Publication Highlighted On The Front Page of The 2020 MDPI Infrastructures Journal.
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • GeoPandas
 • Big data management (SQL, SQLite)
 • SHAP (SHapley Additive exPlanations)
 • XGBoost (eXtreme Gradient Boosting)
 • LDA (Linear Discriminate Analysis)
 • Web Scraping (Beautiful Soup 4, Selenium, Scrapy, lxml, Requests, Urllib, Octoparse)

Link to Journal Publication Close Window

Machine Learning

Contextual Informed Multi-Sequence to Single-Sequence (CIMS2SS) Model

WORK OVERVIEW:
This study aims to combine data from Georgia Department of Transportation's Continuous Count Stations (CCS) and Video Detection System (VDS) stations. CCS are sparsely located and provide statewide vehicle count and classification data. VDS stations are more densely deployed for monitoring traffic and providing real-time information. By matching VDS stations to CCS sites, a deep learning model called Contextual Informed Multi-Sequence to Single-Sequence (CIMS2SS) is created to generate or estimate CCS data from VDS data. The model uses spatiotemporal contextual information, historical median trends, and spatial distances of VDS stations to predict traffic volume at CCS sites. This can be used for quality control, generating data from VDS where CCS sites are absent, or substituting CCS data when stations are inactive or under maintenance.
 
HIGHLIGHTS:
 • Developed and Delivered Effective Application
 • Published Novel Unrolled Recurrent Transformer Architecture
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Graph Neural Networks
 • Application Development
 • Perceiver IO Modeling

GitHub Repo Close Window

Machine Learning

Improving Crash Classification Modeling with Resampling Methods

RESEARCH OVERVIEW:
This study explores the viability of various resampling methods across both machine learning and classical statistical models.
 
HIGHLIGHTS:
 • First Author
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Resampling Methods (SMOTE, ADASYN)
 • Mixed Logit
 • Nested Logit
 • Biogeme Python package
 • XGBoost (eXtreme Gradient Boosting)

Link to Research Close Window

Machine Learning with Applications

Image Data

Model Pipeline

RESEARCH OVERVIEW:
Developed a machine learning model pipeline for detecting wet pavement conditions from live scenes of traffic cameras. Traffic cameras are readily available throughout the United States. This model used visual characteristics such as road surface reflectance to obtain a 0.97 ROC-AUC score on our prediction set.
 
HIGHLIGHTS:
 • First Author
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • CatBoost by Yandex
 • Nested Cross Validation
 • XGBoost (eXtreme Gradient Boosting)
 • Instance Segmentation (Detectron2-Model-Zoo, Panoptic, ResNet, Mask R-CNN, Coco-Instance-Segmentation)

Link to Publication Close Window

Machine Learning Application to Weigh in Motion Sites

Model Pipeline

RESEARCH OVERVIEW:
This project aims to subcategorize vehicles into thirteen FHWA classes. Featuring an object detraction network differentiate five vehicle types. Wheel cluster data is then used in an auxiliary model to differentiate the remaining classes. The model also, segments and reverse image searches truck logos to mine more comprehensive freight data.
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Representation Learning
 • SCAN (Semantic Clustering by Adopting Nearest Neighbors)
 • Perspective Transformation
 • K-Nearest Neighbors
 • Custom Object Detection (YOLOv4)
 • Object Tracking (DeepSORT)

Subproject Github Repo 1. Subproject Github Repo 2. Close Window

Big Data

RESEARCH OVERVIEW:
This project features a hierarchical model architecture consists of a customized variational autoencoder (VAE) and a time-sequence model. The VAE model targets traffic data anomalies at the daily level while the time-sequence model focuses on anomaly detection at a finer temporal scale (i.e., in five-minute intervals). Based on the results, our VAE model achieved 94.1 percent accuracy on the test dataset. For the time-sequence model, four modern model structures were evaluated, including vanilla recurrent neural networks (RNN), gated recurrent units (GRU), long short-term memory (LSTM) units, and Liquid Time Constant (LTC) networks. The selected LTC network significantly outperformed other time-sequence models with a maintaining a percent root mean squared error of 3.45.
 

 

 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Variational Autoencoders
 • Liquid Time-Constant Networks
 • Neural ODEs
 • Apache Hadoop
 • Apache Spark
 • SQL Database/Coding
 • Latent Semantic Indexing
 • Natural Language Processing (Natural Language ToolKit, Bag of Words, Bag of Words)
 • Web Scraping (Beautiful Soup 4, Selenium, Scrapy, lxml, Requests, Urllib, Octoparse)

Close Window

ASME Publication

RESEARCH OVERVIEW:
Developed a cost-effective ways to quantitatively measure the in-vivo mechanical properties of the Achilles tendon.
 
HIGHLIGHTS:
 • Presented research at ASME IMECE2019
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • MATLAB
 • Cross-Correlation Signal Processing

Link to Publication Full Text Close Window

Deep Learning

RESEARCH OVERVIEW:
This GDOT research project implements deeplearning to monitor roadway level of service (LOS).
 
DEVELOPED PROFICIENCY OF SKILL AND KNOWLEDGE:
 • Perspective Transformation
 • Custom Object Detection (YOLOv4)
 • Object Tracking (DeepSORT)

Close Window

GitHub

Link to GitHub Close Window