Abstract
The Artificial Neural Networks (ANNs) are based on the behavior of the brain. So, they can be considered as intelligent systems. In this way, the ANNs are constructed according to a brain, including its main part: the neurons. Moreover, they are connected in order to interact each other to acquire the followed intelligence. And finally, as any brain, it needs having memory, which is achieved in this model with their weights. So, starting from this point of view of the ANNs, we can affirm that these systems are able to learn difficult tasks. In this article, the task to learn is to distinguish between different kinds of traffic signs. Moreover, this ANN learning must be done for traffic signs that are not in perfect conditions. So, the learning must be robust against several problems like rotation, translation or even vandalism. In order to achieve this objective, an intelligent extraction of information from the images is done. This stage is very important because it improves the performance of the ANN in this task.
TopBackground
The Traffic Sign Classification (TSC) problem has been studied many times in the literature. This problem is solved in (Perez, 2002, Escalera, 2004) using the correlation between the traffic sign and each element of a database, which involves large computational cost. In (Hsu, 2001), Matching Pursuit (MP) is applied in two stages: training and testing. The training stage finds a set of the best MP filters for each traffic sign, while the testing one projects the unknown traffic sign to different MP filters to find the best match. This method also implies large computational cost, especially when the number of elements grows up. In recent works (Escalera, 2003, Vicen, 2005a, Vicen, 2005b), the use of ANNs is studied. The first one studies the combination of the Adaptive Resonance Theory with ANNs. It is applied to the whole image, where many traffic signs can exist, which involves that the ANN complexity must be very high to recognize all the possible signs. In the last works, the TSC is constructed using a preprocessing stage before the ANN, which involves a computational cost reduction in the classifier.
TSCs are usually composed by two specific stages: the detection of traffic signs in a video sequence or image and their classification. In this work we pay special attention to the classification stage. The performance of these stages highly depends on lighting conditions of the scene and the state of the traffic sign due to deterioration, vandalism, rotation, translation or inclination. Moreover, its perfect position is perpendicular to the trajectory of the vehicle, however many times it is not like that. Problems related to the traffic sign size are of special interest too. Although the size is normalized, we can find signs of different ones, because the distance between the camera and the sign is variable. So, the classification of a traffic sign in this environment is not easy.
Key Terms in this Chapter
Information Extraction: Obtention of the relevant aspects contained in data. It is commonly used to reduce the input space of a classifier.
Detection: The perception that something has occurred or some state exists.
Classificatio n: The act of distributing things into classes or categories of the same type.
Pattern: Observation vector that for its relevance is considered as an important example of the input space.
Preprocessing: Operation or set of operations applied to a signal in order to improve some aspects of it.
Artificial Neural Networks (ANNs): A network of many simple processors (“units” or “neurons”) that imitates a biological neural network. The units are connected by unidirectional communication channels, which carry numeric data. Neural networks can be trained to find nonlinear relationships in data, and are used in applications such as robotics, speech recognition, signal processing or medical diagnosis
Backpropagation algorithm: Learning algorithm of ANNs, based on minimizing the error obtained from the comparison between the ANN outputs after the application of a set of network inputs and the desired outputs. The update of the weights is done according to the gradient of the error function evaluated in the point of the input space that indicates the input to the ANN