Semi-supervised Labeling of Vehicle Traffic Dataset Using Active Learning for Traffic Analysis Applications

This repository contains all the references and files for this project. YOLOv8 model training was done using Google Colab's cloud computing service, with an NVIDIA T4 as the selected GPU.

Pre-requisites

The following github repositories are used in this project. Download/install these repositories for future implementations/recreations.

• Ultralytics YOLOv8 - used for feature extraction, inference, detection, tracking, and automated labeling. Can be installed through the pip module pip install ultralytics https://github.com/ultralytics/ultralytics

• Sci-kitlearn - used for data clustering. Can be installed through the pip module pip install sklearn https://github.com/scikit-learn/scikit-learn

• TrackEval - used for evaluation of detection and tracking performance. Download the repository in the link https://github.com/scikit-learn/scikit-learn

Folders and Files

• /158,000 NCTS Images Version 1 contains the 158,000 image daset created by Maclang, et. al.
• /Training Images contains the image datasets selected for training the models C and N. Non-clustered 30k contains the images used for model N, while Clustered 22k contains the images used for model C.
• /Modifications contains the edits to the source code of the relevant repositories used in the project. More details can be found in the folder's readme file.
• /Scripts contains the python scripts used as tools to aid in the automation of project tasks.
• /YOLO Training contains all the training sets and validation sets used in YOLO training, as well as training results.
• /docs contains the presentations and documentations as required by the EEE 196 and EEE 199 courses.
• all_models.zip contains the 7 models developed in this project. The model weights are named depending on the dataset it was trained on:
  • base.pt - initial 3,000 images
  • n1.pt - initial 3,000 images + 1 round of non-clustered active learning (500 images)
  • n2.pt - initial 3,000 images + 2 rounds of non-clustered active learning (1000 images)
  • n3.pt - initial 3,000 images + 3 rounds of non-clustered active learning (1500 images)
  • c1.pt - initial 3,000 images + 1 round of clustered active learning (500 images)
  • c2.pt - initial 3,000 images + 2 rounds of clustered active learning (1000 images)
  • c3.pt - initial 3,000 images + 3 rounds of clustered active learning (1500 images)

Reference links

• For official documentation on YOLOv8 and ultralytics components, go here https://docs.ultralytics.com/#where-to-start
• For community discussions on YOLOv8 implementations, go here or on similar forums https://github.com/orgs/ultralytics/discussions
• CVAT annotation tool https://www.cvat.ai/
• Previous study by Maclang et. al used as reference https://gitlab.eee.upd.edu.ph/miguel.lorenzo.orante/video-dataset-labeling-using-active-learning-with-applications-in-vehicle-classification-and-traffic-flow-rate-measurement