Source-linked AI summary
Real time Detection of Lane Markers in Urban Streets
Mohamed Aly
TL;DR
Urban streets pose difficult lane-detection conditions that highway-focused methods do not fully address. The paper combines top-view transformation, selective Gaussian filtering, RANSAC line fitting, Bezier-spline refinement, and post-processing; it reports 50 Hz operation, comparable current-lane results, and good detection of all visible lane boundaries.
Problem
Urban lane-marker detection must handle vehicles, shadows, poor markings, sharp curves, unusual lane shapes, merges, glare, writings, varied pavement, and slopes.
Method
The approach generates an inverse-perspective top view, filters and thresholds it, fits initial lines with Hough and RANSAC methods, refines them with Bezier-spline RANSAC, and post-processes the detections.
Results
50 Hz operation is reported on 640x480 images, with comparable results for current-lane boundaries and good results for detecting all visible lane boundaries.
Takeaways & Limitations
The system detects all lanes in still images of urban streets across varying conditions, extending beyond detection of only the current lane boundaries.
Takeaways & Limitations
False positives occur around stop lines, crosswalks, passing cars, curbs, and confusing street writings, while tracking and color-based classification remain planned improvements.
Abstract
from arXiv · showhide
We present a robust and real time approach to lane marker detection in urban streets. It is based on generating a top view of the road, filtering using selective oriented Gaussian filters, using RANSAC line fitting to give initial guesses to a new and fast RANSAC algorithm for fitting Bezier Splines, which is then followed by a post-processing step. Our algorithm can detect all lanes in still images of the street in various conditions, while operating at a rate of 50 Hz and achieving comparable results to previous techniques.
I. INTRODUCTION
The paper targets lane-marker detection in difficult urban streets with a fast pipeline that transforms road images into a top view and fits lane boundaries with lines and Bezier splines. It evaluates detection across challenging conditions and supports detecting multiple visible lane boundaries.
- Urban lane detection is harder than highway detection because streets contain vehicles, shadows, poor markings, sharp curves, unusual shapes, lane merges, glare, road writings, varied pavement, and slopes.
- The pipeline applies inverse perspective mapping, selective Gaussian filtering, thresholding, simplified Hough detection, RANSAC line fitting, RANSAC spline fitting, and post-processing.
- 50 Hz on 640x480 images is reported for the real-time system on an Intel Core2 2.4 GHz machine.
- The system detects all visible lane boundaries, including neighboring lanes, rather than only the current lane.
- Inverse perspective mapping removes perspective convergence so lanes become vertical and parallel, while restricting processing to a subregion reduces runtime.
- The IPM transformation uses camera intrinsics, pitch, yaw, and camera height under a flat-road assumption to map image points onto the road plane.
B. Filtering and Thresholding
The IPM image is filtered with an oriented, separable Gaussian kernel tuned to lane geometry, then thresholded by retaining only the strongest responses. This emphasizes bright, near-vertical lane markers while enabling efficient processing.
- The vertical filter component is a smoothing Gaussian whose σ_y corresponds to a 1 m lane-segment height in the IPM image.
- The horizontal component is a second-derivative Gaussian whose σ_x is tuned to an expected lane width of 3 inches.
- The filter targets vertical bright lines on a dark background but can also respond to quasi-vertical lines.
- The separable kernel supports faster implementation than a non-separable kernel and produces high responses at lane markers.
- 97.5% is the experimental quantile threshold, retaining only the highest responses from the filtered image.
C. Line Detection
Line detection first counts and localizes candidate vertical lines in the thresholded image, then robustly fits them with RANSAC.
- C. Line Detection: A simplified Hough transform sums thresholded pixel values by column, smooths the sums, detects local maxima, and refines their positions to sub-pixel accuracy.Nearby lines are then grouped.
- C. Line Detection: RANSAC line fitting operates within windows around each detected vertical line to robustly fit the candidate lines.The windows are illustrated in Figure 6.
- C. Line Detection: The resulting RANSAC step produces fitted lines from the localized windows.Figure 6 shows the fitted lines on the sample image.
D. RANSAC Spline Fitting
The method refines candidate lines with a fast RANSAC procedure for fitting third-degree Bezier splines, scoring candidates efficiently to favor longer and straighter lane boundaries.
- D. RANSAC Spline Fitting: Candidate lines initialize a third-degree Bezier spline fitting stage that can refine straight-line guesses into curved lane detections.The spline’s control points form a bounding polygon around the spline.
- D. RANSAC Spline Fitting: RANSAC spline fitting repeatedly samples points, fits a Bezier spline by least squares, computes its score, and retains the best-scoring spline.The algorithm is organized around three functions inside the main loop.
- D. RANSAC Spline Fitting: Sample parameters t_i ∈[0, 1] are assigned proportional to cumulative Euclidean distance from each sampled point to the first point, with t_1 = 0 and t_n = 1.This associates the first and last sampled points with the spline endpoints.
- D. RANSAC Spline Fitting: The fitted control points minimize the sum of squared errors for the sampled points using a pseudo-inverse solution.The matrix P is solved from the sampled-point system.
- D. RANSAC Spline Fitting: Instead of computing point-to-spline distances through a fifth-degree equation, the method rasterizes each spline and counts image-pixel values to compute its score.The score is designed for efficient RANSAC evaluation.
- D. RANSAC Spline Fitting: The scoring formula favors longer and straighter splines by penalizing them less than shorter and curvier splines.It uses normalized length and curveness measures with regularization factors.
E. Post-processing
Post-processing improves spline localization by searching across the spline’s normal direction and refitting with localized image evidence.
- E. Post-processing: The algorithm samples points along each initial spline, searches normal line segments for smoothed grayscale maxima, and refits the spline using the localized points.This is performed in both the IPM image and the original image after back-projection.
1) Localization:
After improving spline position, the method extends splines beyond their endpoints by searching along tangent directions and accepting new normal-direction image peaks.
- 1) Localization:: Post-processing produces splines that appear longer and more localized on the lanes.Figure 10 compares blue pre-processing splines with green post-processing splines.
- 1) Localization:: Spline extension searches forward and backward from the endpoints along tangent directions, using normal line segments to locate additional grayscale peaks.The extension is performed in both IPM and original image spaces.
2) Extension:
Geometrical checks validate localized and extended splines, replacing or rejecting fits that violate curvature, length, or orientation constraints.
- Localized splines that are very curved or very short are replaced by the corresponding RANSAC line fit.
- Fitted splines are rejected when they are not near vertical in the IPM image.
A. Setup
The evaluation uses four hand-labeled urban-street clips covering varied conditions and compares detected lane boundaries with ground truth using spline-distance criteria.
- 1224 labeled frames containing 4172 marked lanes form the hand-labeled evaluation dataset.
- The clips include curves, street writings, differing pavement types, direct sunlight, shadows, and passing vehicles.
- Each detected lane boundary is automatically compared with ground-truth lanes to classify correct and false detections.
- Spline identity is assessed by sampling points on each spline and computing nearest-point distances in both directions.
- The method uses median and mean bidirectional distances with experimental thresholds t1 = 20 and t2 = 15.
B. Results
The algorithm performs effectively across varied urban conditions, supports both current-lane and all-lanes detection, but produces false positives in specific visual contexts.
- The algorithm achieves comparable current-lane results to methods that use both detection and tracking, despite processing images independently.
- All-lanes mode detects visible lane boundaries beyond the current lane and achieves good results on this previously unattempted task.
- False positives arise from stop lines, crosswalks, passing cars, and curbs mistaken for unpainted right lane boundaries.
- Color classification and temporal tracking are proposed to reduce false positives from curbs and yellow street writings.
V. CONCLUSION
The paper presents a real-time urban lane-detection pipeline that detects all lanes in still images and operates at 50 Hz, with robustness illustrated across challenging conditions.
- False detections include confusion with street writings, crosswalks, curbs, vehicles, and stop lines on cross streets.
- The pipeline combines a road top view, Gaussian filtering, line detection, RANSAC spline fitting, and post-processing.
- The method detects all visible lanes in still urban-street images under shadows, vehicles, curves, varied road structures, and pavements.