An Analysis of the Road Signs Classification Based on the Higher-Order Singular Value Decomposition of the Deformable Pattern Tensors - Advanced Concepts for Intelligent Vision Systems - page 198

different signs. Some signs are also very similar, especially if geometrically
changed, e.g. 30 km/h compared to the 80 km/h, etc.
As alluded to previously, the method tolerates well imperfections in detection
(such as not well aligned window, etc.), as well as variations in color tint and/or
lighting obtained in real road conditions (we used the Marlin® and Sony® cameras).
The method is also resistant to the slight projective deformations (i.e. allowed for the

Fig. 5. Examples of correct classification of the signs despite the imprecise detection and under
different color and/or lighting conditions. In all cases the color images were converted to the
monochrome versions by taking exclusively the blue channel. Then intensity values were
conditioned by histogram equalization method.

Fig. 6. First five tensors T for the 40 km/h speed limit sign (upper row), and the
corresponding five core tensors

Z (lower row)

road signs in respect to the drivers' direction of view), as well as to some occlusions if
these do not affect the main part of the pictogram. Correct operations under different
operating conditions are presented in Fig. 4 and Fig. 5, for instance.
Training of the data base in Fig. 8a takes around 8-9s in our platform, while run
time classification is in order of 0.04-0.07s per single image of resolution 640 480,

depending on a size of the test pattern (the difference in computation time depends on
time necessary for the geometrical registration to the test pattern).

×