[TORCS] End-to-end learning for highway assistance driving system

In the previous post (2017/07/28 - [TORCS] - End-to-end learning for autonomous driving), I trained a Convolutional neural network that produces a steering wheel angle command from a front-facing camera input image. 

In this post, I will show you a self-driving car that keeps the current lane position while keeping a constant distance from the closet car ahead.  

The input to the CNN is a (normalized RGB) front-view camera image of size (130 x 320) pixels. The output of the CNN is a vector of size (2 x 1). The first element of the output vector is the distance from the closest car ahead. The second element is the steering wheel angle command.

The training data is obtained from TORCS while driving tracks in TORCS. (For more details, see 2017/07/28 - [TORCS] - End-to-end learning for autonomous driving). About 150,000 samples are acquired from 10 tracks in TORCS. 

The spec. of the CNN is similar to VGG19 except for fully connected layers. The weights of the CNN are initialized by Xavier method and for the error back-propagation, Adam optimizer is utilized.

  

(This post will be updated soon.)