Automatic Traffic Sign Recognition

Nowadays a growing number of sources is producing complex digital images; the need for object detection arises also in the automotive world. In particular driver's assistance systems aim at a continuous analysis of the road looking for obstacles, lane marks and traffic signs. In this research the aim is to automatically detect the traffic signs along the road based on their shape, irrespective of brightness, position, size, orientation. To this end theoretical tools from Image Analysis, Mathematical Morphology, Pattern Recognition and Machine Learning are linked together to form a modular detection framework. Images are first decomposed by looking at connected components of constant intensity and the Max-Tree is set up; the Max-Tree is interpreted as a tree of nested peak components from which the shape features are extracted by means of image moments. The moments are collected and used to train supervised classifiers as CCM, kNN and LVQ. In order to have a useful training data set a labeling step turns out to be necessary; the labeling is performed by using a self-made application realized during a precedent stage. The labeling stage led to the construction of a labeled data set (ground truth); during this phase the limits of a naive reference-based filtering became apparent and the usefulness of contrast-based filtering emerges very clearly. Later also a preprocessing step turned out to be necessary; this because of the huge number of peak components deriving from each of the gray images. Preprocessing revealed difficulties in the data set reduction by considering the peak component area and the moment similarity to ideal traffic sign and results in sampling as viable solution. The use of different Minkowsky metric looks marginal as well. Training the classifiers on the sampled data set led to other insights. First the problem seems to be learnable by kNN and LVQ but not by CCM; this confirms preprocessing difficulties and indicates the distribution of the classes are not hyperspherical and well separated. The performances of kNN reveal the added value of considering higher order moments. The LVQ classifier obtains good performances after a key statistical data reshaping and reveals limits deriving from the skewed distribution of the traffic sign examples. One-class classification aspects implied by the background patterns emerge as well; also the use of several prototypes hits cross-validation limits given by this dataset. A proof-of-concept application is realized demonstrating real-time detection of traffic signs in digital images of urban roads.