A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer...

11
A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008

Transcript of A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer...

Page 1: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

A Robust Method for Lane Tracking Using RANSAC

James Ian VaughnDaniel Gicklhorn

CS664 Computer VisionCornell University

Spring 2008

Page 2: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Original Project Plan

• Break into slices to both allow for non-linear solutions as wellas reducing time for line fitting.

• Detect predominant edgemarker or road edgeif no markers present

• Algorithm must be robust enough to handle when markingsare not present or badly defined.

Project Plan

Page 3: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Road Model Assumptions

• Road edges are strong

• Road edges have a gradient similar to the model edge

• Road edges are consistent between frames

• Roads consist of several segments of approximately linear edge points

• Actual edges have a lot of inlier points

Model Assumptions

Page 4: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

RANSAC as a Line Fitting Algorithm

• Fast efficient line fitting

• Dynamic selection ofcandidate lines withtunable error constraint

• Able to incorporate spatialpriors and progressivemodel fitting into constraint.

• Can detect predominant road edges along side if lanemarkers aren’t present.

RANSAC

Page 5: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Implementation

• Using RANSAC, we can incorporatea progressive spatial prior intothe error function.

• Each frame is broken in to road“slices”, consistent, horizontal sections most likely to contain road markings.

• This reduces the amount of time needed for RANSAC as well asproducing a non-linear output

Implementation

Page 6: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Candidate Error Criteria

• The RANAC error function is the heart of our method.

• The error considers both local andpriori models to locate “good”line candidates in the scene

• Lines are restricted by slope,distance from a priori “model”line, and by the edge intensity,density, and gradient angle oftheir potential “inlier” members.

RANSAC Error Model

Line End 1

Line End 2

Distance 1

Distance 2

Threshold

Dis

tanc

e Fr

om T

est

Mod

el

Model Line

Test Line

angle

AngleMag

natu

de

angl

e

Test Point

Page 7: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Candidate Error Criteria

RANSAC Error Model

Line End 1

Line End 2

Distance 1

Distance 2

Threshold

Dis

tanc

e Fr

om T

est

Mod

el

Model Line

Test Line

angle

AngleMag

natu

de

angl

e

Test Point

Error is a linear combination of:• Scaled distance of candidate line to model line• Angular difference between candidate and

model lines.• Number of consensus points to a given candidate.• Gradient angle difference between local points

to a candidate line and the model line.• Density of consensus points to a candidate line.• Gradient magnitudes of local points to a candidate line.

Page 8: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Candidate Error Criteria

RANSAC Error Model

Line End 1

Line End 2

Distance 1

Distance 2

Threshold

Dis

tanc

e Fr

om T

est

Mod

el

Model Line

Test Line

angle

AngleMag

natu

de

angl

e

Test Point

Error is a linear combination of:• Scaled distance of candidate line to model line

- Maintains spatial / temporal consistency.• Angular difference between candidate and

model lines.

- Maintains spatial /temporal angular consistency.

• Number of consensus points to a given candidate.

- Looks for best probability of a good line.• Gradient angle difference between local points to a

candidate line and the model line.

- Looks for consistent gradients to model line.• Density of consensus points to a candidate line.

- Prefers denser line edges to lines.• Gradient magnitudes of local points to a candidate line.

- Prefers most pronounced lines.

Page 9: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Results

Simulation Results• Results show excellent marking

detection and outlier tolerance.

• With proper calibration andweighting, model performswell.

• Outputs show need fortemporal filtering, butoutliers are withintolerance levels.

Page 10: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Results

Videos

Page 11: A Robust Method for Lane Tracking Using RANSAC James Ian Vaughn Daniel Gicklhorn CS664 Computer Vision Cornell University Spring 2008.

Conclusions

Conclusions and Future work• RANSAC is an easy to constrain / almost modular method

for line detection and does well in scaling to the scene where others can fail.

• Future work to improve this model would include• Better optical flow• Robust calibration• Optimal parameter tuning• Temporal Filtering• Vehicle Sensor Integration