Programming_An_Autonomous_Car

Table of Contents

1. Overview
2. Challenge Details
   1. Training data
   2. Challanges
   3. Equipment
   4. Deploying Models to the Car
   5. Training Machines
   6. Tracks
   7. Driving Scenarios

Overview

In this project the task is to train a deep learning algorithm to autonomously navigate a real car around a realistic test circuit, and make the appropriate manoeuvres where necessary. At the end of the project, you are expected to give a presentation and write a report about what you have done. Your model will be tested on the track and will compete against the models of your peers.

Challenge Details

• Work in pairs
• Develop a deep learning model
• Input = Image from a camera on the car
• Predictions = Appropriate speed and steering angle

Training data

• Dataset is hosted on Kaggle
• 13.8k images
• Spped & Steering angle is also available in the dataset
• We are free to generate our own dataset

Challanges

1. This will allow you to automate the process of model submission, and obtain an indication of performance (using a small set of test data), before we evaluate them on the final, unseen data.
2. Kaggle competition is hosted here.
3. Create a Kaggle account (if you do not have one) and form a team with your project partner.
4. A live challenge, where your pre-trained model will be deployed to the car and tested on real circuits. This will be performed in person.

Equipment

• The main body of the car is the SunFounder PiCar-V kit V2 and is equipped with a Raspberry Pi (RPi)
• TensorFlow v2.4 is installed on the car,
• The car has an optional Coral Edge TPU, which is a custom device to run forward-pass operations for edge computing.
• Note that it isn’t necessary to convert your model to TensorFlow Lite.

Deploying Models to the Car

A standardised skeleton code will be provided to you that you should integrate your pre-trained model with, which we will then install on the car prior to the live testing.

Training Machines

• We can use Google Colab or our own local machine to train the model
• We will also have access to MLiS1 or MLiS2 machines (each with w) to perform training. These are accessible by ssh’ing into the machine, by typing
  • ssh username@mlis1.nottingham.ac.uk or
  • ssh username@mlis2.nottingham.ac.uk, where username is your University username.
• In order to install custom packages on your machine, you will need to set up a conda environment. To install conda, type the following command
  • bash /shared/Anaconda3-2019.10-Linux-x86 64.sh
  • Once installed, you will need to add a start up script
  • echo . ∼/.bashrc >> .profile
  • Lastly to create your conda environment use
  • conda create --name my env python=3.6

Tracks

• T-junction tracks
• Oval Track
• Figure-of-eight track.

Driving Scenarios

Important: we only use UK driving rules, i.e. driving on the left-hand side. The training data was based on the following driving scenarios:

1. Keeping in lane driving along the straight section of the T-junction track.
2. As (1), but stopping if a pedestrian is in the road.
3. As (1), but driving as normal if pedestrians or other objects are on the side of (but not in) the road.
4. Driving around the oval track in both directions.
5. As (4), but stopping if a pedestrian is in the road.
6. As (4), but driving as normal if pedestrians or other objects are on the side of (but not in) the road.
7. Performing a turn at the T-junction, in response to a traffic sign (either left or right).
8. Driving around the figure-of-eight track in both directions, continuing straight at the intersection. We will not consider objects in or at the side of the road for this scenario.
9. Stopping at a red traffic light and continuing at a green traffic light. We will only consider these scenarios in the live testing.

Files in the Repository

1. angle.ipynb calculates the angle while driving through the different tracks
2. speed.ipynb calculates the speed while driving determining when to stop and accelerate