2. DATA
COLLECTION
Simulating Self driving car on a
video game. Initially user runs
through the road using his/her
mouse and collects two kinds of
data:
• IMAGE (shot from camera
mounted on car)
• STEERING ANGLE
3. DATA COLLECTION
DETAILS
1. Input Data: Image of road captured by
car’s camera every second.
(Background of the slide is Input image)
2. Labeled Data: Steering Angle – A numeric
number between -10 to +10 stored in csv file
after every second.
• Steering Angle = 0 (drive straight)
• Steering Angle = Negative (drive left)
,High Negative Value – indicates steep
curve.
• Steering Angle = Positive (drive right)
,High Negative Value – indicates steep
curve
(Background image’s top left shows negative
number)
4. APPROACH : NEURAL NETS
Supervised Learning
Use of CNN (Convolution Neural
Network)
• Many existing models tested like
GoogleNet, VGNet.
• Best Model on which it worked – Nvidia
Self Driving car Model (CNN based).
• https://arxiv.org/pdf/1604.07316v1.pdf
5. Before Training on CNN
Car crashes –
obvious due to
absence of
trained model.
(Click on video to
run it)
6. Trained Model (CNN) - Half Track Covered
• Since we have a trained model on
CNN, it runs well most of the track.
• One possible reason - Unbalanced
data set as 80% of track has straight
road thus our data has 80% of zero
values in steering angle (labeled
data set).
Road Type Car is able to run
?(Yes/No)
Straight Yes
Non- Steep Curve Yes
Steep Curves No
7. Trained Model – Full Track Covered
• In previous slide, car crashed during
curve which comes at 13 second.
• In current slide, car is able to move
through the curve as model was
improved due to – more data
collection around curves,
optimization of parameters, etc.
• Shortcoming of current model : See
Next Slide.
8. Overfitting
• If we observe car did crossed the full lap but its
movement was more zig zag as compared to
video in slide # 6. Reason is over-fitting.
• Maybe a perfect example to show where over-
fitting makes business’s product unviable as no
one would like to sit in a zig zag car .
• Training self driving car is strenuous , requires
GPU’s and many more parameter optimization.
• That’s one possible improvement as part of this
project.