vikasnataraja / Using-Machine-Learning-to-Switch-Lanes-in-Autonomous-Cars

This project explores the use of machine learning applied to image data from a single camera to decide switching lanes. Dataset is from UC Berkeley's DeepDrive group.

Using-Machine-Learning-to-Switch-Lanes-in-Autonomous-Cars

This project explores the use of machine learning applied to image data from a single camera to decide switching lanes. Dataset is from UC Berkeley's DeepDrive group.

What this project is about:

This project explores the use of ML in autonomous cars when only a single camera is available and no other data is. Essentially, how does machine learning work when the data is limited to a single camera?

How it works:

We are using the data from UC Berkeley's DeepDrive research group. They have over 100,000 images and videos from autonomous cars. Unfortunately, the actual dataset itself is a highly nested .json file (COCO format). It has various features including multiple object detections, lane detections, weather and more. You can read more in their research paper.

The full dataset is a 1.4GB .json file that is available here.

Since the dataset was complicated, we decided to use pandas to simplify it to a dataframe. You can read our detailed report to know exactly how we did the feature engineering. We also decided to choose 1300 images to hand-label the outcome variable i.e. 1 meaning can change lanes, 0 meaning do not change lanes. That 1300 image dataset is here.

If you want to see the images, then go to http://bdd-data.berkeley.edu/ where you can download them. If you want to see the code for how we did the feature engineering and formatting, you could do that here.

The final result of the formatDataset.py is a .csv file

Using the csv file and the outcome variable, we used various machine learning models to calculate the accuracies.

What you will need to run this:

• Python 3
• Scikit-Learn
• Pandas package

How to run:

Easy way

• Navigate to main.py and using a python interpreter, run the program (the files X and y must be in the same directory)

The not-so-easy way

• If you want to start the dataset construction from scratch, which I highly recommend you don't, go to formatDataset.py and run.
• This uses the finaltrain1.json file which has data from1300 images. I highly recommend starting with a smaller dataset like example_3.json which has data from only 2 images and then making your way up to 1300.

References

The dataset, the images, the labels, the research paper are all from UC Berkeley. Check their website, it's pretty cool stuff.

Co-authors

Claire Savard, CU Boulder

Vandana Sridhar, CU Boulder