The tasks of autonomous rescue robots operating in unknown environments are manifold. Self localization, map generation and the detection of possible victims are indespensable. Apart from these, other factors can become crucial for the survival of the involved persons and for the safe operation of the robot itself. A first step in autonomously detecting such dangers is the real-time recognition of standardized danger signs in camera images. The knowledge of such information can be incorporated into the exploration algorithm as well as enhance the generated maps for later usage by human rescue teams. Our approach is a combination of histogram backprojection and speeded up robust feature (SURF) matching. The first one is used to detect regions of interest within the image. In the second step, interest points are extracted and their features are calculated. These features are then matched against the samples in a database, taking into account the constraints resulting from the affine transformation of the matching objects. We have tested the approach on a set of 240 scene images containing 5 different kinds of hazard signs. In 90 images, none of the signs was present, but objects of similar size and color. The approach detected 92% of the signs if the signs filled at least five percent of the pixels in the 1024 times 768 pixels image. None of the fake objects were detected as a hazmat sign in these experiments. The approach was implemented and successfully tested in practice on our mobile system "Robbie X", which was used by the team "resko" at the RoboCup World Championship 2008 in Suzhou, China.