Local Image Features Extraction for Road Signs

Question

I'm working on a project to detect a road sign from the Image. Consider the sample image below:

I'm using a heuristic search to detect the road sign within an image. Basically, I want some weak features (colour, shape, zero symmetry within road signs) which can be calculated very fast and their weighted average can be used recognize the road sign.

Currently, I'm searching the entire image to find the road sign which could be of different scale. Since the domain of search is a lot, I tried to reduce it bu by using some heuristics:

1. Location of sign detect in previous frame (its mostly likely to be near)
2. Road signs are usually on right sign
3. Road sign contains enough white pixels

But eventually search needs to stop when it finds the region that contains a road sign. For this purpose, I'm relying on some fast and robust local features that predicts likelihood of containing road sign within given region of an image. Currently, I'm relying mainly on colour and some region props that Matlab provides.

I tried looking for lines and corners within edge map of the region even tried thresholding and looking for rectangle using morphology. But seems like they are too binary either its YES or its NO. In fact in most cases it fails, so my questions what are some nice local features that I could rely on to detect road sign?

I would really appreciate if somebody you guide in the right direction or provide some links or papers that could be helpful for this problem.

Answer 1

Do you know what signs you are looking for? If yes, maybe you could do template matching (e.g. a normalized cross-correlation, available in matlab). It won't work great when signals are getting closer, since the perspective projection will change their appearance, but it should work for mid-range detection.

You can limit your template-matching search to the right side of the road.

There are also template matching algorithms for affine transformations that might be a next step.

Christos

Answer 2

You could take a look at recent publications by Segvic et al, I know they have been working on the problems of traffic sign detection.

The basic idea was to use the Viola-Jones framework for object detection, which was later improved by adding some temporal and spatial constraints. If I remember correctly, they achieved a nearly 100% recall rate with just 2 false positives on 10000 frames (I might be wrong with the numbers, but they are something like that).

The best is to refer to the papers directly (sorted by year):

Sinisa Segvic, Karla Brkic, Zoran Kalafatic, Vladimir Stanisavljevic, Marko Sevrovic, Damir Budimir and Ivan Dadic. A computer vision assisted geoinformation inventory for traffic infrastructure. ITSC, Madeira, Portugal, September 2010: 66-73

Sinisa Segvic, Karla Brkic, Zoran Kalafatic, Axel Pinz. Exploiting temporal and spatial constraints in traffic sign detection from a moving vehicle. Machine Vision and Applications (accepted for publication)

Sinisa Segvic, Zoran Kalafatic, Ivan Kovacek. Sliding Window Object Detection without Spatial Clustering of Raw Detection Responses. IFAC SYROCO 2012

In case I forgot to link something relevant, see the other publications here.

Answer 3

I think something worthy for you to have a try is the pixel co-occurrence model. Roughly speaking, this is something based on pixel intensities. So you can quickly skip all unlikely positions and concentrate on those interested ones. Search for articles about skin models.

Meanwhile, there are some HAAR orthogonal projection based recognition algorithms (this idea is published on CVPR and PAMI). Although I have not used it before, it is not surprising that this algorithm is quite fast, because its computations are integral image-like. In other words, you can match an arbitrary size of rectangular ROI in a target image.