Tracking Occluded Lane-Markings for Lateral Vehicle Guidance

Jean-Philippe Tarel, Frederic Guichard and Didier Aubert

Abstract

---

We present a real time algorithm for computing the orientation and the lateral pose of a vehicle with respect to the road observed by an on-board video camera. The advantage of this approach is to provide robust measures when lane-markings are dash, partially missing, perturbed by shadows, highlights, other vehicles, and noise. Moreover, a calibrated camera may be used as well a uncalibrated camera.
Robustness to intensity perturbations is obtained, as much as possible, by taking into account all the edges in each image without doing any a priori thresholding based on the gray-level amplitudes. This leads to numerous edges to be processed. Nevertheless, we propose an algorithm extracting straight line segments directly in gray-level images in less than 0.05 second for a 256x256 image on a Pentium 200Mhz.
For robust estimates of the orientation and of the lateral pose under geometric perturbations such as missing data, the algorithm relies on few basic assumptions on the observed cues: aligned and straight cues. Moreover, we assume that the road profile is changing slowly from one frame to another.
The orientation of the vehicle with respect to the road is computed by estimating the position of the focus points of the markers along the line of horizon. From the extracted straight line segments, the algorithm builds an histogram which represents the current lateral road profile. The relative lateral pose of the vehicle is obtained by comparing the lateral profile of the current road to a reference lateral road profile. Using the consistency over time of the lateral road profile, the reference is updated. Initial reference is built during the initialization of the process when the vehicle is assumed well enough aligned with the road.
In the proposed approach, if lane-markings are missing, estimated orientation and pose can still be computed using the other cues seen in the scene such as cues from side-walks, road shoulders, herb-sides, and guard rails. This allows us lateral guidance without explicit recognition of the left and right lane-markings as it is usually proposed.
First experiments on real images show correct estimations in presence of dash lane-markings, missing markers, highlights, shadows, curves, and noise. The whole process can run typically at a rate from 5 to 15 images per second, and the obtained measures can be used for helping the driver in the lateral guidance of its vehicle, in case of uncontrolled lane departure for instance. This system may help toward a solution of major security problems of road traffic, such as obstacle detection near the vehicle.

Reference

---

Pdf file (722 Kb)

(c) WSEAS