I've attached a picture with my question. In this picture there are some lines to train the neural network to learn the path of the road.
I was wondering is there an specific neural network for that like YOLO or Faster RCNN?
or can I just add some outputs for the points of the lines?
Mohammad Hasani ,
I would suggest that before choosing a particular NN be sure to have all that is required to solve the problem. For example, what is the frame rate that you are going to feed the NN the pictures? what is the image resolution? What is the speed of the vehicle? How are you going to use bounding boxes ?
These are important questions to answer before choosing an NN architecture. For example, to see how the resolution is a factor see:
https://www.researchgate.net/post/What_Image_processing_techniques_can_be_used_to_effectively_detect_railway_track_lines?
A NN is a pattern classifier and you need to give it focus, too little and you loose information, too much (as in the link provided ) and you risk loosing the NN's power (and processing) due to irrelevant details.
If you put more details on implementation it would help in narrowing down the possibilities and assumptions that will be made (or have to assume).
You could start with line detection methods.
For the NN, do you plan to detect vehicles as other objects or only lanes?
Well the answer is not simple, but mainly depends on what you want to achive. The key tech to this are deep learning networks. Path detection from 2D pictures might not work in real time applications like in a car, because it can easily be fooled and is more or less an edge detection algorithm in the 2D picture. doing some preprocessing in advance. Doing a working path detection is more complex and probably also involes an extra dimension "time". So typically you use a high frequency of 2D frames probably 10-30 frames per second and try to analyze how the picture change over time. practically you need hundred thousend of sequences to train and test the network. As you want to do supervised training I guess, you have to prepare this sequences properly for the objects, that you want to identify. Modern consumer photo cameras have this kind of algorithms for auto focus tracking built in as they use 2D, depth and time (sequence) to follow faces or objects. In your case your object to follow is the road track.
I hope that gives you an idea where to search further.
Kind regards, Wolfram
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training