Lanes Detection Based on Unsupervised and Adaptive Classifier Andrés F. Cela Department of Automation and Industrial Control Escuela Politécnica Nacional Quito, Ecuador andres_cela@ieee.org Franklin L. Sánchez Department of Electronics Instituto Tecnológico Superior Sucre Quito, Ecuador fsanchez@tecnologicosucre.edu.ec Luis M. Bergasa Department of Electronics University of Alcalá Alcalá de Henares, España bergasa@depeca.uah.es Marco A. Herrera Center for Automation and Robotics Universidad Politécnica de Madrid Madrid, Spain marco.herrera@ieee.org Abstract— This paper describes an algorithm to detect the road lanes based on an unsupervised and adaptive classifier. We have selected this classifier because in the road we do not know the parameters of lanes, although we know that lanes are there, only they need to be classified. First of all, we tested and measured the brightness of the lanes of the road in many videos. Generally, the lines on the road are white. We used the HSV image and we improved the region of study. Then, we used a Hough transform which yields a set of possible lines. These lines have to be classified. The classifier starts with initial parameters because we suppose that the vehicle is on road and in the center of the lane. There are two classes, the first one is the left road line and the second one is the right road line. Each line has two parameters that are: middle point of line and the line slope. These parameters will be changing in order to adjust to the real lanes. A tensor holds the two lines, so these lines will not separate more than the tensor allows. A Kalman filter estimates the new class’s parameters and improves the tracking of the lanes. Finally, we use a mask in order to highlight the lane and show to the user a better image. Keywords- Lanes detection, lanes classifiers, Kalman filter. I. INTRODUCTION There is a high requirement in the car industries of implementing road lanes identifier systems on board. This is to give more safety to passengers. Because this, the vision- based lane detection has been an interesting research area. Nowadays, most of applications are make off-line. There are many techniques for detecting the road lanes and most of them use vision with mono-camera. Although, using a stereo camera is being common. In this area there are two types of detections, the road detection and the lane road detection. The approaches in these fields in the last 5 years is surveyed in [1]. The DARPA grand challenge 2006 helped to make significant progress for commercial driver’s assistance systems. In this challenge vehicles has a lot of sensors, to detect road, signs, edges, traffic lights, and of course the lane marks. However, researches have detected that the bottleneck in this systems is the road perception [1]. The researches in [2] show how to detect the road region based on homography estimation. They use a set of cameras and using a three module algorithm can detect the road. The work in [3] also uses a stereo camera for detecting the road with the 3D information. Using a stereo camera can give some advantages over a monocular camera, but in normal applications to use a monocular camera is cheaper. On the other hand, it is better to make low cost applications in order to implement them in smart phones, tablets or things like these, which will be our future work. A more simple system is showed in [4]. This is based on a roads classifier which decides if the image is a road or not. This classifier has a data base of many kinds of roads. Also, there are many models and researches to detect the road lanes. Only 8 % of the works are in real time algorithms. It is still complex to implement a real time algorithm, because for obtaining better results good and big processors are necessary. Using the vanishing point is a way to find the lanes, for example [5]. This work has 94% of true detections, but it does not work in real time. When a lane is not uniform or is a curve, it is more difficult to use the Hough transformation or the basic RANSAC. In [6] is presented a solution for curve lanes and in [7] a solution for heterogeneous lanes. The work in [8] is based on RANSAC to get the road lanes. This work uses a special camera which is able to collect 500 frames per second. This is used for high speed vehicles. To detect the road lanes is necessary to consider the environmental conditions, like the light, rain, brightness, etc. For example, the researches in [9] present a solution for night conditions. This is based on the light reflections of the lines of the road. Then, a Support Vector Machine (SVM) classifies these lines. In some cases it is necessary to know the Euler angles of the camera, like in [2], [5] y [8], but in this work we do not use these angles. First of all, these angles have to be 2013 Fifth International Conference on Computational Intelligence, Communication Systems and Networks 978-0-7695-5042-8/13 $26.00 © 2013 IEEE DOI 10.1109/CICSYN.2013.40 228