Skip to main content

DriveSafe v1

Binary Classification of Safe and Risky Driving

The first version of DriveSafe was very crude. I found this pyimagesearch article, and thought it was acceptable to use the code as a base to bootstrap the project.

I did not have a very complete understanding of every line of the code, however I was able to grasp the high-level concepts behind it at the time. Though I just wanted to get something working, this was crucial to the success of the project.

Categories​

  • v1 was very basic and had only a binary feature set: safe and risky.

Demo​

In this demo, I had set the prediction averaging algorithm deque to size of 1 (meaning that there was no flickering-averaging of predictions). This was because I wanted to see the raw output of the model.

Improvements​

  • You might already have noticed in the demo that due to the nature of the model not capturing the temporal features of the data, the predictions are very jittery and not very reliable.
  • Especially for features such as lane weaving, it would be hard for even a human to determine if the driver is weaving lanes or not, let alone a deep-learning model that is trained on relatively scarce data.
  • The solution: On further research, I had discovered the LSTM layer (which is a type of RNN) that could be used in conjuction with the ResNet50 CNN to capture both the temporal and image features of the data.

Training Plot​

training_plot

Hyperparameters​

  • Batch Size: 32
  • Epochs: 25
opt = SGD(learning_rate=1e-4, momentum=0.9, decay=1e-4 / args["epochs"])
model.compile(loss="categorical_crossentropy", optimizer=opt,
	metrics=["accuracy"])

print("[INFO] training head...")
H = model.fit(
	x=trainAug.flow(trainX, trainY, batch_size=32),
	steps_per_epoch=len(trainX) // 32,
	validation_data=valAug.flow(testX, testY),
	validation_steps=len(testX) // 32,
	epochs=args["epochs"])

Dataset​

  • This was a very simple model that was trained on a very small dataset of 688 images (risky) and 700 images (safe).
  • The rationale behind selecting this number of images (around 700 per category) is that I was following as closely to the pyimagesearch article as I could, and this was the number of images they used in their dataset.
  • https://www.kaggle.com/datasets/voidranjer/drivesafe/data

Model​

Code​

Footnotes​

Some important changes that were made to the code (from the pyimagesearch article):