Abstract
A prerequisite for any system that enhances drivers' awareness of road conditions and threatening situations is the correct sensing of the road geometry and the vehicle's relative pose with respect to the lane despite shadows and occlusions. In this paper we propose an approach for lane segmentation and tracking that is robust to varying shadows and occlusions. The approach involves color-based clustering, the use of MSAC for outlier removal and curvature estimation, and also the tracking of lane boundaries. Lane boundaries are modeled as planar curves residing in 3D-space using an inverse perspective mapping, instead of the traditional tracking of lanes in the image space, i.e., the segmented lane boundary points are 3D points in a coordinate frame fixed to the vehicle that have a depth component and belong to a plane tangent to the vehicle's wheels, rather than 2D points in the image space without depth information. The measurement noise and disturbances due to vehicle vibrations are reduced using an extended Kalman filter that involves a 6-DOF motion model for the vehicle, as well as measurements about the road's banking and slope angles. Additional contributions of the paper include: (i) the comparison of textural features obtained from a bank of Gabor filters and from a GMRF model; and (ii) the experimental validation of the quadratic and cubic approximations to the clothoid model for the lane boundaries. The results show that the proposed approach performs better than the traditional gradient-based approach under different levels of difficulty caused by shadows and occlusions.
Highlights
According to the World Health Organization (WHO), about 1.2 million people get killed in traffic accidents each year worldwide, while the number of injured is estimated to be 50 million
The accuracy and precision of the lane detection and tracking algorithm is measured in terms of the mean absolute error (MAE), the root mean square error (RMSE) and the standard deviation of the error between the estimated lane geometry and the manually identified curves for the lane boundaries in the analyzed datasets
The MAE and RMSE errors are computed pointwise, i.e., if the ground truth curve is represented by a set of points (xi, y(xi )), i = 1, 2, . . . , N, sampled from the real lane boundary, and the estimated curve is given by (xi, ŷ(xi )), i = 1, 2, . . . , N, the estimation error at point xi is ei = y(xi ) − ŷ(xi )
Summary
According to the World Health Organization (WHO), about 1.2 million people get killed in traffic accidents each year worldwide, while the number of injured is estimated to be 50 million. We propose a vision-based lane detection and tracking method capable of estimating the lane geometry and relative vehicle position even if lane boundaries are not clear due to shadows, changes in illumination or partial occlusions caused by other vehicles. The proposed lane sensing system should help to enhance the safety of drivers and pedestrians by preventing unintended lane changes due to distracted driving or reducing risky maneuvers due to excessive speed for a given lane curvature Another contribution of this paper is the comparison of the textural features considered, which were generated with two textural models: (i) Gabor features; (ii) a Gaussian Markov random field model. The proposed approach relies on well-known mathematical tools or models, some of which have been employed in previous work in the field of lane detection, and this article should have a reference and tutorial value
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
