Ensemble of ANN for Traffic Sign Recognition

Abstract

“Machine Learning (ML) is the subfield of Artificial Intelligence conceived with the bold objective to develop computational methods that would implement various forms of learning, in particular mechanisms capable of inducing knowledge form examples or data” (Kubat, Bratko & Michalski, 1998, p. 3). The simplest and best-understood ML task is known as supervised learning. In supervised learning, each example consists of a vector of features (x) and a class (y). The goal of the learning algorithm is, given a set of examples and their classes, find a function, f, that can be applied to assign the correct class to new examples. When the function f takes values from a discrete set of classes {C1, .…., CK,}, f is called a classifier (Dietterich, 2002). In the last decades it has been proved that learning tasks in which the unknown function f takes more than two values (multi-class learning problems) the better approach is to decompose the problem into multiple two-class classification problems (Ou & Murphey, 2007) (Dietterich, & Bakiri, 1995) (Massulli & Valentini, 2000). This article describes the implementation of a system whose main task is to classify prohibition road signs into several categories. In order to reduce the learning problem complexity and to improve the classification performance, the system is composed by a collection (ensemble) of independent binary classifiers. In the proposed approach, each binary classifier is a singleoutput neural network (NN) trained to distinguish a particular road sign kind from the others. The proposed system is a part of a Driver Support System (DSS) supported by the Spanish Government under project TRA2004-07441-C03-C02. For this reason, one of the main system requirements is that it should be implemented in hardware in order to use it aboard a vehicle for real time categorization. In order to fulfill this constraint, a reduction in the number of features that describe the instances must be performed. As consequence if we have k generic road sign types we will use k binary NN and k feature selection process will be executed.

Key Terms in this Chapter

Artificial Neural Network: Structure composes of a group of interconnected artificial neurons or units. The objective of a NN is to transform the inputs into meaningful outputs.

(*)Feature Selection: Process, commonly used in machine learning, of identifying and removing as much of the irrelevant and redundant information as possible.

Weka: Collection of machine learning algorithms for solving data mining problems implemented in Java and open sourced under the GPL.

Feature Space: n-dimensional space where each example (pattern) is represented as a point. The dimension of this space is equal to the number of features used to describe the patterns.

One Against All: Approach to solve multi-class classification problems which creates one binary problem for each of the K classes. The classifier for class i is trained to distinguish examples in class i from all other examples.

Field Programmable Array (FPGA): A FPGA is an integrated circuit that can be programmed in the field after manufacture.

Correlation-based Feature Selection: Feature Selection(*) algorithm which heuristic measures the correlation between attributes and rewards those feature subsets in which each feature is highly correlated with the class and uncorrelated with other subset features.

Machine Learning: Computer Scientific field focused on the design, analysis and implementation, of algorithms that learn from experience.

K-Cross-Validation: Method to estimate the accuracy of a classifier system. In this approach, the dataset, D, is randomly split into K mutually exclusive subsets (folds) of equal size (D1, D2, …, Dk) and K classifiers are built. The i-th classifier is trained on the union of all Dj ¤ j¹i and tested on Di. The estimate accuracy is the overall number of correct classifications divided by the number of instances in the dataset.