Self-Driving Picar

Code for a Raspberry Pi Model B self-driving car (heavily modified version of picar).

How it works

Components

• Raspberry Pi Model B
• GETIHU Power Bank BG-120
• HP Webcam HD-4110
• Ovonic 11.1V 2200mAh Lipo Battery (yes, it's overkill)
• L298N H-bridge motor controller
• 4-Wheel Drive Yellow Robot Smart Car Chassis

Artificial Intelligence

The car uses image recognition to determine which direction the car should go in (left, right, or straight). It is built to stay on a track made with two lines of black duct tape.

To teach the car to drive, I captured 20,000 images of me controlling the car with my phone. From this data, I trained a deep neural network with Keras.

Getting started

Prerequisites

Install Node.js on your Raspberry Pi

sudo apt-get install -y nodejs

Install node modules on your Raspberry Pi

npm install zerorpc socket.io node-static

Install python libraries on your Raspberry Pi

pip3 install zerorpc opencv-python tensorflow numpy

Install Python (with pip) on your Windows PC

Install python libraries on your Windows PC

pip install opencv-python tensorflow numpy

Setting up the car

1. Build a car using the same (or similar) components
2. Set the pi's hostname in hostname.txt
3. Move files from pi_code to your Raspberry Pi (under Documents/sdc/)

.\upload_to_pi.bat

4. Make the scripts executable

chmod u+x manual.sh auto.sh record.sh

Usage

Driving

Run the following script on the Pi to drive in "manual" mode

./manual.sh

or in "auto" mode

./auto.sh

or in "record" mode

./record.sh

Then navigate to [raspberry-pi-ip-address]:8080/socket.html on your phone's web browser.

Example link: http://192.168.1.106:8080/socket.html

Wait a few seconds for everything to initialize... and enjoy!

Recording your own training and testing data

                                                            File to run
1. Modify 'constants.txt' if necessary                  |   ---------------
2. Clear 'train_data/' of all files                     |   ---------------
3. Record raw data on pi                                |   record.sh
4. Download raw data from pi                            |   update_data.bat
5. Process data                                         |   process_data.py
6. Add processed data to training data                  |   merge_data.py
7. Repeat steps 3-6 until training data is full         |   ---------------
8. Duplicate and flip training data                     |   flip.py
9. Repeat steps 2-8, but this time for test data        |   ---------------

Training your own neural network

1. Download and unzip the dataset or use your own
2. Put training data under train_data and testing data under test_data
3. Edit the hyperparameters in train.py
4. Train the neural network: python train.py

License

MIT