Lane Detection Based On Feature Extraction Methods

A genuine challenge of the automotive industry is to design an economical and trustworthy Advanced Driver Assistance System(ADAS).ADAS is a system that provides assistance to drivers while driving.It can effectively assist the driver,reduce the number of traffic accidents,and enhance road safety and passengers' comfort.With the development of autonomous vehicles,lane detection is one of the key technologies of ADAS.This study proposes a lane detection algorithm based on the combination of several feature extraction methods.When lane markers are examined,typical lane markers on the road comprise reflective paints that are yellow or white,which generate their own distinctive shapes.The prime goal is to use these characteristics of lanes to propose an adaptable ROI and algorithm for lane detection in both image and video.The proposed method comprises three major modules: initialization,lane detection,and lane tracking.The initialization stage focuses on extracting features to select a proper Region of Interest(ROI)in the frame.The initialization stage comprises sky region detection,region boundary,and selecting ROI respectively.The sky region detection module detects the sky region area by using gradient information and energy function.The module gives a horizon line and feeds into the region boundary module with it.The horizon lines distinguish the sky and non-sky area on the frame.The region boundary module extracts color features by using HSV color space and extracts objects' regional boundaries applying to Felzenwalb's method.Both color and regional boundaries feature frames that are separately applied to Hough Transform to verify that the candidate lines are in both feature frames.Selecting the ROI module determines the proper region of interest using the horizon line and the candidate lines from the region boundary module.It scans extensive areas so as not to miss candidate lane markers on the road.After determining ROI,ROI is transformed into a Birdseye view using the IPM method to remove perspective effects.A scan-line method in the lane detection module applies to the Birdseye view to ascertain the lanes and their positions.Finally,a Kalman filter in the lane tracking module applies to optimize and predict the posterior location of the lane markers based on the previous location of the lane markers.The proposed method does not require any pre-defined region of interest while each module in the initialization feeds the next module to determine a region of interest.The combination of several features makes the proposed method robust.The efficiency of the proposed method is evaluated on the Caltech benchmark dataset and one custom video sequence.Moreover,experimental results show that the proposed method works well in various road conditions and achieves real-time performance in video processing.Innovation of the proposed method can approximately calculate the curvature of the road and estimate road types such as two-lane roads,and solid lines.The detection ratio and execution time for the proposed algorithm in both datasets are compared with those of other studies,and it is shown that the proposed algorithm is better.