@ Mohamed Abdel-Nasser Thank you for your reply, I attached below an example of a road scene in a synthetic image. As you can see, we have different types obstacles on the scene; we have cars and mountains. Mountains are irrelevant in my application. My task is to detect the closest obstacle in the scene, I am using Matlab to sketch my method wich will be implemented on FPGA platform for real time performance. Is there a simple way to detect sideway pixels, I really know a very little about image classification and I can't use any classifier here because real time performance must be achieved. Thank you very much

I am assuming that this is a first person view of the scene(the picture represents the actual view and it is not a third person view for the car in front).

You can use a conditional probabilistic model where you initialize the conditional model based on a window that is currently on the bottom row of pixels and assume that that is the road, then move up the scene and calculate the probability P(y_i+1|y_i) and you can do so using a context window on the x dimension also(so that you can infer where does the edge of the road is). The algorithm should do this within frame and across frames to update the meaning of the assumption of the road.

I have laid out the bare minimum, but in principle based on the information within frames and across frames you can potentially infer:

• The obstacle or bifurcation ahead(the road continues on both sides but in the middle the probability will be low). Both instances can be treated the same if you are going to pass the car in front of you
• edges of the road
• Turns such as the one in the picture where you detect an almost zero probability in the y dimension(road does not continue forward, but there is a continuation on the x dimension )

I have also assumed that illumination conditions vary slowly with respect to the sampling rate so that it can be assumed stationary

This is also a very rough sketch of the idea which needs further computational details but can get you started.

Arturo Geigel Thank you for you response, I understand the concept of conditional probability. and I understant from your response that I can use a conditonal model to ditinguish between pixels that belong to the road and those that do not. Unfortunately, I am not well experienced with how to use such models with image processing, can you please give me some refrences to read so that I can understand well what you are suggesting

The closest that I could find to guide you is:

http://users.utcluj.ro/~vision/index_files/Publications/ITSC2004.pdf

Though slightly different from my approach it has enough of the details that can be adapted to the method I gave you or you could try the method in the paper.

Regards