Behavioral Cloning Project

Overview

---

(Vehicle driving autonomously)

In this project, a convolutional neural network is trained to clone a driving behavior from a simulator. The CNN architecture used here is referenced from NVIDIA's End to End Learning for Self-Driving Cars paper.
Reference: https://arxiv.org/pdf/1604.07316v1.pdf

This repository contains follwoing five files:

• model.py (script to create and train the model)
• drive.py (script to drive the car - feel free to modify this file)
• model.h5 (a trained Keras model)
• a report writeup file (the document you are reading right now)
• video.mp4 (a video recording of the vehicle driving autonomously around the track for one full lap)

Environment

---

you need the simulator.

I try to run on my locally end, but there are too many problem for version, so I used the workpace on udacity.

The Goal of the Project

The goal is to create and to validate of a CNN model that clones a driving behavior from a simulator.
Simulator allows to collect data in the format shown below:

Here, only 3 camera images (center, left, right) and corresponding steering angles are used to train the model.
Thus, the outcome of the model is the steering angle.

Data Collection Strategy

---

The data is collected via the simulator.
While collecting the data, only 1 loop has been conducted as follows:

1. One round of data was collected by staying in the middle line as much as possible and making smooth turns
2. Totally use 1694*3 (center, left and right) for training.

The Model

---

1. An appropriate model architecture has been employed

My architecture is using Keras Sequential() method and base on NVIDIA's End to End Learning for Self-Driving Cars paper.
Reference: https://arxiv.org/pdf/1604.07316v1.pdf

---

2. Attempts to reduce overfitting in the model

The model contains dropout layers in order to reduce overfitting.
The model was trained and validated on different data sets to ensure that the model was not overfitting.
The model was using 6th epochs.

---

3. Model parameter tuning

The model used an adam optimizer, so the learning rate was not tuned manually.
The model used Dropout rate was 0.5.
The model used epochs rate was 6.
The Architecture

---

4. Appropriate training data

Training data was chosen to keep the vehicle driving on the road. I used a combination of center lane driving, recovering from the left and right sides of the road.
I am not a good driver. I try my best to keep the car on the center of road.
I totally use 4611 record to train.

---

5. The architecture

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
cropping2d (Cropping2D)      (None, 90, 320, 3)        0         
_________________________________________________________________
lambda (Lambda)              (None, 200, 66, 3)        0         
_________________________________________________________________
lambda_1 (Lambda)            (None, 200, 66, 3)        0         
_________________________________________________________________
conv2d (Conv2D)              (None, 98, 31, 24)        1824      
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 47, 14, 36)        21636     
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 22, 5, 48)         43248     
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 20, 3, 64)         27712     
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 18, 1, 64)         36928     
_________________________________________________________________
dropout (Dropout)            (None, 18, 1, 64)         0         
_________________________________________________________________
flatten (Flatten)            (None, 1152)              0         
_________________________________________________________________
dense (Dense)                (None, 100)               115300    
_________________________________________________________________
dense_1 (Dense)              (None, 50)                5050      
_________________________________________________________________
dense_2 (Dense)              (None, 10)                510       
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 11        
=================================================================
Total params: 252,219
Trainable params: 252,219
Non-trainable params: 0

Result

I try to be simple,so only trained one lap.

see the result gif

Suggestions

1.Environment-Because I can't run on local end, so I can't improve my model freguently. I can get better model on local end.
2.Improve model- The model architecture is base on NVIDIA's. It has some space to improve.
3.Data clear- some data maybe are bad(bad-control,wrong-control) which need delete to improve model.