self-driving-car-simulator

The core technology behind Self Driving Cars today. Given the image of a road at a time frame, it can decide where to turn the steering and how much. I am working continuously to generalize it to as many different terrains as possible.
It uses a Convolutional Neural Network to predict the motion of the steering given the image of a road at a time.

Requirements: Python 3.5 ,Keras 2.0.2 , Tensorflow 1.2.1 , OpenCV 3.2, numpy 1.11.0

This approach uses Regression for predicting the angle of steering, and is clearly more successful and accurate than the classification approach which I used before. Regression provides flexibility to the results.

Regression approach

Classification approach

How to:

Use:

 $ python drive.py   
 You can use it either on a live video feed from the webcam, or a pre saved video on disk for the demo.    

Train your own model:

 Use the script cnn_train.py or train_cnn2.py(branch 2 named Regression-Approach). Make sure the datset is ready.

Generate the dataset:

 Use the script generate_data.py to generate the dataset.    
 It requires the path of a video on disk from which training samples will be generated along with the action taken by the user.    
 It automatically puts a frame in the right folder(class) according to actions taken by user while generating data.     

Contents /Scripts:

-cnn_train.py :

    To train the  model.    

-train_cnn2.py:

    This file is in branch named "Regression-Approach". This is to train the regressive model.    

-generate_data.py :

    To generate the dataset from random videos.    

- simulator_gui.py :

    The class that provides the GUI for simulator.    

-drive.py

    The main script that starts the simulator.    

-model2.json, model4.json :

    The pre trained models on 4 differnt terrains. Note that the model2.json is different in both the branches.    

-weights2.hdf5, weights4.hdf5 :

    Weights of the corresponding models.    

About The Model:

The classification based model:

The regression based model:

Trained using Backpropogation algorithm with stochastic gradint descent.

Accuracies after 10 epochs:

For classification based model:

-Train acc: 96.4665%    
-Test acc : 88.5039%     

It may seem like it has been overfit. But no. It was the test set, which contained some wrong examples.

For regression based model:

-Train error: 2.0311  (Mean absolute error)     
-Test error:  2.4532