In this notebook, I classify traffic images as 'accident', 'dense_traffic', 'fire', or 'sparse_traffic.'
The data can be found here: https://github.com/OlafenwaMoses/Traffic-Net. It consists of 4400 .jpg images of traffic, evenly distributed among the four classes.
I use a number of techniques to classify the images, including a simple dense network, a convolutional neural network, and transfer learning. I also experiment with data augmentation and dropout.
Using accuracy to evaluate my models, I achieve the following results:
| Model | Accuracy |
|---|---|
| Baseline (random) | .25 |
| Dense Net | .4 |
| CNN | .67 |
| CNN with dropout & data augmentation | .78 |
To code this notebook, I relied heavily on the instruction provided in the book Deep Learning with Python, Second Edition by François Chollet.