Could anyone suggest the best algorithm for a real-time Car recognition (say in a parking space)? I am planning to implement the same on FPGA as well. Kindly suggest. Thank you.
$\begingroup$
$\endgroup$
2
-
1$\begingroup$ Do you have any pictures of the parking lot? This can make a big difference on your recognition method. My Initial guess though is having areas of interest marked (each parking stall) and then doing background subtraction $\endgroup$– andrewCommented Feb 25, 2015 at 19:44
-
$\begingroup$ Currently planning on taking Satellite Images. $\endgroup$– KrishCommented Feb 25, 2015 at 23:37
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
1
Do you mean car detection or recognition? Detection: where are all the cars? Recognition: which car is this? See: Object detection versus object recognition
In any case, the short form of the answer is to prototype the heck out of this, likely using OpenCV or libccv, your favorite neural network package, and as much horsepower as possible (several Nvidia GTX 980). Don't worry about realtime-ness or implementation details, just get it working first. The state of the art for object recognition and detection both is probably convolutional neural networks, check out stuff like AlexNet; there are simpler algorithms (eg Histogram of Oriented Gradient or HoG) which could do okay at detection, but not at recognition (there is a decent person detector using HoG, I think you can probably make a decent car detector in the same way).
Running this on a FPGA is... unlikely: high-end desktop GPUs are very well suited to the problem, and orders of magnitude faster at it. People are working on neural network ASICs now, should be ready in a couple of years.
-
$\begingroup$ Recognition. For example, I want to count the number of cars of each model, i.e.: Tesla - 2 Mercedes - 3 etc. $\endgroup$– KrishCommented Feb 26, 2015 at 18:41
5
-
$\begingroup$ Okay. How about the others such as SIFT, SURF? Does it also help? Or is ISM more accurate or something? $\endgroup$– KrishCommented Feb 25, 2015 at 23:38
-
$\begingroup$ Well, they could find a single, or a couple of objects. Finding a full set of database is more cumbersome. So in a way, yes ISM is more accurate. The code is also available: code.google.com/p/object-detection-implicit-shape-model/source/… $\endgroup$ Commented Feb 26, 2015 at 0:26
-
$\begingroup$ Thanks! I will have a look and try it out. Thanks again. $\endgroup$– KrishCommented Feb 26, 2015 at 0:28
-
$\begingroup$ And this is in OpenCV if I am not wrong! $\endgroup$– KrishCommented Feb 26, 2015 at 0:28
-
$\begingroup$ @Krish Please read - dsp.stackexchange.com/help/someone-answers. Citation - "Please do not add a comment on your question or on an answer to say "Thank you". Comments are meant for requesting clarification, leaving constructive criticism, or adding relevant but minor additional information – not for socializing. If you want to say "thank you," vote on or accept that person's answer, or simply pay it forward by providing a great answer to someone else's question." $\endgroup$– SergVCommented Feb 26, 2015 at 4:01