Vishnu Karthik
Pronouns: He/Him/His
Research Mentor(s): Stuart Batterman, Professor
Research Mentor School/College/Department: Environmental Health & Safety, School of Public Health
Presentation Date: Thursday, April 22, 2021
Session: Session 2 (11am – 11:50am)
Breakout Room: Room 18
Presenter: 5
Abstract
Automated vehicle detection and counting systems are becoming increasingly valuable to monitoring highway data. However, few systems have implemented such algorithms for use with a dashcam setup. Based on a review of traditional traffic monitoring methods, a You Only Look Once (YOLO) v3 Convolutional Neural Network (CNN) was trained via Darknet using scraped images of vehicles and traffic from Google. Then, every vehicle within each image was labelled by vehicle type (Personal, Commercial or Heavy Duty) to train the model. The loss curve generated shows a healthy learning rate for the model and successful classification of vehicles for images under ideal conditions. The results indicate that the YOLOv3 CNN works well for classifying vehicles from dashcam photos. Further research is needed to identify factors that could strengthen the performance in low-light and complex background/scenery environments.
Authors: Vishnu Karthik, Stuart Batterman
Research Method: Laboratory Research
