traffic-ai

Set Up

pip3 install -r requirements.txt

The data set (gtsrb folder) can be downloaded here.

Running

$ python traffic.py gtsrb

Video

Findings

In all my attempts I kept the compilation parameters constant.

model.compile(
    optimizer="adam",
    loss="categorical_crossentropy",
    metrics=["accuracy"]
)

At first I attempted a model using the ReLu activation function and a hidden layer of 64 nodes, once again with the ReLu activation function. The accuracy was only around 10-15%.

Then I removed the hidden layer altogether and used a softmax activation function. The accuracy went up to around 80%.

I then added a hidden layer with 8 nodes and a softmax activation function. Immediately, the accuracy dropped to 5%. Same thing once I increased the number of nodes to NUM_CATEGORIES

Then I replaced the hidden layer by a convolutional layer, which improved the accuracy to 90%. I also added a max pooling layer which increased the accuracy to 93%.

Doubling the number of filters on the convolution layer didn’t significantly improve the accuracy.

I changed the convolutional layer’s activation function to sigmoid, which increased the accuracy to 98%

I added a hidden layer 8 * NUM_CATEGORIES nodes but the accuracy actually dropped to 94%

I found that increasing other parameters to get accuracy over 97-98% dramatically decreased performance so the cost seemed higher than the improvement.

I also added dropout.

Project Details