The focus of this research is on designing deep learning based tools for automated highway condition assessment. At this stage, we primarily focus on developing holistic algorithmic constructs for fully autonomous detection and labeling of road asset items. To this end, we have developed a CNN-based deep classifier based on VGG-Net at front-end of our proposed framework. At the same time, the deep classifier takes the benefits of transfer learning to overcome the challenge of limited data with high sparsity, and binary networks for finetuning of assets classification at the back-end of the framework.
In addition to python 3.5, you also need to install the follwoing python packages:
Tensorflow, pandas, SciPy, scikit-learn, seaborn, Six, PIL, Image, matplotlib, opencv-pythonResize all the images into a unique size by using imresize.py. Make sure to have all images under a folder called imgs in the same directory.
Save the data in a folder named as dataset in two sepearate subfolders as following in a categorical fashion (One folder for each class):
train_dir = './dataset/train'
valid_dir = './dataset/validation'Apply the required changes based on the size of input images, number of images, etc. Then run train.py.
Note: alex-train.py is a code that we wrote to try using AlexNet for this project before we started using transfer learning on VGGNet.
Once training and test is done, use the script misclass-analysis.py for misclassification analysis. Save the misclassification raw data extracted from train.py in a CSV file named classifier_results.csv as the input of this script. Run the code in Jupyter Notebook to see the flow of result generation.
Sadegh Nouri Gooshki - [email protected]
Copyright (c) 2018, University of North Carolina at Charlotte All rights reserved. - see the LICENSE
We would like to thank the Virginia Department of Transportation (VDOT) and Leidos for providing the image dataset.