Manuscript - CIARP 2014 - 19th Iberoamerican Congress on Pattern Recognition A fast pavement location approach for autonomous car navigation Thiago Rateke, Karla Justen, Vito Chiarella, Rodrigo Linhares, Antonio Carlos Sobieranski, Eros Comunello, Aldo von Wangenheim INCoD – Brazilian Institute for Digital Convergence www.incod.ufsc.br Abstract. This paper describes a fast image segmentation approach de- signed for pavement detection in a moving camera. The method is based on a graph-oriented segmentation approach where gradient information is used temporally as a system of discontinuities to control merging be- tween adjacent regions. The method presumes that the navigable path usually is located at specific positions on the scene, and a predefined set of seed points is used to locate the region of interest. The obtained results shown the proposed approach is able to accurately detect in an in- expensive computation manner the navigable path even in non-optimum scenarios such as miss-painted or unpaved dirt roads. Validation was con- ducted using a dataset with 701 samples of navigable paths, presenting a very high precision for real-time applications. 1 Introduction Image segmentation has been extensively studied over the past decades in many application areas. Its goal is to divide an input image I into a set of homogeneous regions W , where each W i represents a particular object on the scene. This process can be performed by grouping similar patterns based on some similarity measure, such as vector norm [1], statistical metrics [2], specific color metrics (HSV, CIE-lab) for better color discrimination, and others. Once obtained, these regions can be used as primary information in high-level algorithms specifically designed for object recognition. Specifically for the road segmentation problem, three main approaches can be found in the literature: feature-based, model-based and lane-region detection [3]. The feature-based approaches employ of some kind of markers to identify lane boundaries, typically extracted from the scene with gradient information proceeded by Hough transform [4]. Methods based on this approach can produce excellent lane detection results [5], but they are highly dependent on an optimum scenario to be applied [6]. In a non-optimum scenario where the lanes are not entirely painted, methods from this category tend to produce unsatisfactory results [6]. On the other hand the model-based detection approaches try to use a robust system to predict the road shape from a straight line or a parabolic curve [3][6]. Since this approach is less dependent on the lane markers it normally produces more satisfactory results than the feature-based algorithms, but at some extra computational cost. And even though they may take into account complex roads, as S curves [7], they do not recognize bifurcations, an important requirement for a completely autonomous navigation. The last category, the lane-region detection approaches, attempt to tag each pixel as road or non-road,