Zahra Rahimi Afzal et al. International Journal of Recent Research Aspects ISSN: 2349-7688, Vol. 7, Issue 2, June 2020, pp. 51-58 © 2020 IJRAA All Rights Reserved page - 51- Lane Departure Detection Using Geometrical and Intensity Patterns Zahra Rahimi Afzal 1 , Behzad Nazari 2 , Saeid Sadri 3 1 Computer Science Department, Kansas State University, zrahimi@ksu.edu 2 Electrical and Computer Engineering Department, Isfahan University of Technology, nazari@cc.iut.ac.ir 3 Electrical and Computer Engineering Department, Isfahan University of Technology, sadri@cc.iut.ac.ir Abstract- A lane departure warning system (LDWS) is an essential part of an intelligent transportation system. This paper proposes a novel low-complexity LDWS that detects lane departures in video frames captured by smart phones with various lighting conditions and lane types and complicated road surfaces. The car used in the research was assumed to be traveling on a mostly straight road or highway signed with lane markings, and left and right lane markings were expected in fixed regions of frames. The Canny edge detector detected all the edges, allowing extraction of connected edge components. Left and right lane markings were selected from these components according to the position, orientation, and pixel intensity pattern. The presence or absence of lane markings in some consecutive frames was used to detect lane departure. This algorithm operated in real time and was successfully implemented on a tablet. Keywords— Lane departure detection, Lane departure warning system, Driver assistance, Connected component I. INTRODUCTION Intelligent transportation systems, or driver assistance systems use lane detection to implement lane departure warnings, Lane Keeping Assist technology, lateral control, and collision warnings in order to decrease vehicle accidents and fatalities. A lane departure warning system (LDWS) warns a driver when driver inattention causes unintended lane departure. There have been many researches based on image processing in transportation systems like [1, 2, 3]. For example, [3] uses K- means clustering and Kalman filtering to classify highway lanes. Our application utilizes image processing methods to detect lane departure using a mobile phone mounted behind the windshield of a car. The next following literature review is limited to our application. Vision-based lane detection algorithms can be classified as feature-based, region-based, or model-based [4]. Feature-based methods detect lanes by merging low-level features, such as the edge or color of lane markings. However, feature-based methods are susceptible to existing noise or occlusion. The Generic Obstacle and Lane Detection (GOLD) system in [5] is a stereo vision-based system that recognizes lane structures. The perspective effect of the two stereo images was removed using Inverse Perspective Mapping (IPM), and the resulting image was used to detect lanes with morphological filtering. Region- based methods utilize classification methods to detect lanes in a road. These methods initially extract proper features and then classify image pixels as lane and non-lane. The research in [6] used color and texture features to segment images into road and non-road groups. Model-based methods use a few parameters to represent a geometric model (straight or curved, etc.) for lane markings [7]. Although these models are less susceptible to noise and occlusion than feature-based methods, they contain calculation complexity and they utilize only special forms of road [5]. Wang et al. proposed a B-snake-based flexible lane model that detects complex lane structures, such as S-shape lane markings by the setting of control points [8]. A robust algorithm extracted a proper initial position needed for the geometric model. The suggested method is applicable to any type of marked and unmarked roads with shadows and lighting variations. In [9] a Catmull-Rom spline was proposed for flexible modeling of multilane detection with extended Kalman filter tracking and no assumption regarding the form and parallelism of lane markings. The research in [10] used the IPM algorithm and a bank of steerable filters to separate lane markings with various orientations, and then Random Sample Consensus (RANSAC) algorithm applied a parabolic model to fit the road. Lane model parameters were achieved via Kalman filtering. In [11] the algorithm initially applied selective oriented Gaussian filters on a top view of the road image, and then Bézier splines were used to model lane markings using the RANSAC algorithm. The research in [12] proposed use of a geometric model and Gabor filter to recognize lane markings.