5
$\begingroup$

In the context of autonomous driving, two main stages are typically implemented: an image processing stage and a control stage. The first aims at extracting useful information from the acquired image while the second employs those information to control the vehicle.

As far as concerning the processing stage, semantic segmentation is typically used. The input image is divided in different areas with a specific meaning (road, sky, car etc...). Here is an example of semantic segmentation:

enter image description here

The output of the segmentation stage is very complex. I am trying to understand how this information is typically used in the control stage, and how to use the information on the segmented areas to control the vehicle.

For simplicity, let's just consider a vehicle that has to follow a path.

TL;DR: what are the typical control algorithms for autonomous driving based on semantic segmentation?

$\endgroup$
  • 1
    $\begingroup$ Can you provide more details and add some sources? $\endgroup$ – DuttaA Aug 23 '18 at 13:41
  • 2
    $\begingroup$ Thanks. I added two reference videos and I tried to explain my question better $\endgroup$ – firion Aug 23 '18 at 14:13
  • 1
    $\begingroup$ @DuttaA ,we should always keep this in mind,we arent here to stream video clips,we value time for other global scientific problems. $\endgroup$ – quintumnia Aug 23 '18 at 18:12
  • 2
    $\begingroup$ @quintumnia well I saw the question was downvotd so I thought the question lacked sources...It is not my field so I really don't know what OP posted on these links $\endgroup$ – DuttaA Aug 23 '18 at 18:17
  • 1
    $\begingroup$ I don't understand how the new paragraph could be useful in explaining my question better. It does not add any detail... $\endgroup$ – firion Aug 24 '18 at 6:51
1
$\begingroup$

This is a great question. Connecting the AI knowledge to the application is a difficult task and its not easy to test and perfect.

There are a few points regarding autonomous driving algorithms that one should keep in mind.

  • A lot of data is required before actually applying a model to the real world.
  • The control algorithms are designed for specific tasks, there is no master algorithm that solves the issue.
  • Time is very important, decisions have to be made in real time.

The image segmentation you have mentioned in the question is rarely used in current autonomous technologies. The main issue is that

  • The model is slow and requires massive computing resources. Due to this is cannot help in real time decision making.
  • A very important parameter depth is missing from the information. This is of utmost importance in autonomous driving technology. A 2D image cannot provide this parameter.

So the current technology use

  • A range of sensors (LIDAR, Ultrasonic, Cameras etc.) that provide readings 100+ times a second.
  • A lot of the computer vision algorithms and control algorithms are hard coded. There are set parameters within which decisions are made.
  • The control algorithms are designed to work with each set of sensors and the mechanical components of the car.

To get a better look at the way these algorithms are used, you can have a look at

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.