Udacity Self Driving Car Term1

I have recently finished the Udacity self-driving car course Term 1.

The course is great and out of my expectation. I joined the course because I want to consolidate my understanding of neural network an computer vision. The course does fulfill my wishes.

In term 1, the course teaches you the followings:

Use computer vision to find the road

The driving computer needs to know where the road is in order to drive. This is not a simple task, and if there are errors, the car can screw up heavily.

There are 2 projects about finding roads in videos.

Basic: writeup,

result (1), result (2), result challenge

Advance: writeup, result (1), result(2), result(3)

Use Neural network to detect traffic sign

The driving computer must be able to read traffic signs. There is a paper which explains how to read German traffic sign at an accuracy of 98%, which is must higher than human.

I cannot faithfully reproduce this paper and obtain a test set accuracy of 96.2%.

My network failed to recognize the “Pedestrians” sign, which I can say it is not desirable.

Here is my write-up

Use Neural network to mimic human driving

This is an interesting project, Nvidia team published a paper to train the driving computer to drive around using just videos of human driving around. This assignment is to reproduce the paper in a simulator.

I am quite satisfied with this project, I train the car to drive forward, but in the end, the car can drive backward, which implied that neural network is not just a memorizer, and it does learn something.

Here are my write-up and the results (1), (2), (3), (4)

Use HOG-SVM to detect vehicles on the road

The driving computer must be able to detect vehicles on the road. In the project, I have a thousand of car images and use it to train an SVM machine to decide if an image segment contains a vehicle.

When processing the video, I have to break up the video input into many segments and pass it to SVM, if SVM says it is a car, I mark it on the image and finally create an image where the car is highlighted.

Here is my write-up and result

Conclusion

It pretty much a taste of writing a driving computer. I can say if I load my program and use it to drive a car. The car crashes at no-time. I need to further enhance these components before I can move forward before I can invent my self-driving car (or a self-driving RC car).