Robust Lane Detection and Tracking in Challenging Scenarios
read more
Citations
Data-Driven Intelligent Transportation Systems: A Survey
Recent progress in road and lane detection: a survey
Towards End-to-End Lane Detection: an Instance Segmentation Approach
Perception, Planning, Control, and Coordination for Autonomous Vehicles
Deep Neural Network for Structural Prediction and Lane Detection in Traffic Scene
References
Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography
Multiple view geometry in computer vision
Multiple View Geometry in Computer Vision.
Sequential Monte Carlo methods in practice
Related Papers (5)
Video-based lane estimation and tracking for driver assistance: survey, system, and evaluation
Frequently Asked Questions (17)
Q2. What have the authors stated for future works in "Robust lane detection and tracking in challenging scenarios" ?
Future work will integrate it with a vision-based obstacle-detection algorithm, for example [ 20 ], for a collision-warning system.
Q3. How is the pixel value of the rectified image calculated?
Since u and v are not integer numbers in most cases, the pixel value of the rectified image is calculated by linearlyinterpolating the intensity values of the four neighboring pixels (by flooring and ceiling the u and v) in the original image.
Q4. What is the way to classify a hypothesis?
When there are n features (inputs) and m hidden nodes, it requires nm multiplications, nm + m additions, and m sigmoid-function calculations to classify a hypothesis (n = 27 or 81 and m = 7 in their examples).
Q5. How do the authors normalize the size of the lane marking?
Since the size of the lane marking changes dramatically with respect to its distance from the car, the authors need to normalize them to apply a standard classifier.
Q6. What is the way to improve the detection performance of lane markings?
Applying a stereo algorithm [3], [4] can further improve the lane-marking-detection performance, but the authors focus on a monocular image in this paper.
Q7. Why is it good to give a reasonable weight to this?
when the detection performance is good enough, it is good to give a reasonably large weight to this because redundant detection compensates tracking failures.
Q8. Why is the motion of the lane boundaries in world (vehicle) coordinates?
Due to vehicle’s vibration, including pitch change, the motion of the lane boundaries in world (vehicle) coordinates is not smooth enough to be properly modeled by a Kalman filter.
Q9. How did the authors obtain the ROC curves for all the classifiers?
For all the classifiers, the authors obtained the ROC curves by changing only the threshold values (no relearning with different parameters).
Q10. Why did the authors choose a particle filter over the Kalman filter?
the authors chose a particle-filtering algorithm over the Kalman filter to prevent the result from being biased too much on the predicted vehicle motion but to give more weight to the image evidence.
Q11. Why was the motion of the vehicle modeled by Gaussian distributions?
For the particle filtering, the vehicle’s motion (rotation and translation) was modeled by Gaussian distributions for simplicity, but the scoring function is carefully designed to prevent the result from being dictated by this model.
Q12. What is the way to reduce the computation time?
To further reduce the computation time, the authors applied a cascade classification: First, a simple gradient detector and an intensity-bump detector with loose (low) threshold values are successively applied to quickly filter out nonlane markings, and then, the ANN classifier is applied to the remaining samples (much smaller in number).
Q13. How many hypotheses are selected per lane boundary?
In their implementation, up to five hypotheses per lane boundary (left/right) are selected, including the ones from the particle-filtering process.
Q14. How many control points are generated from a random set of two line segments?
An approximate arc of three control points is generated from a random set of two line segments, and a more complicated hypothesis of four control points is generated from a random set of three line segments.
Q15. What is the way to generate a straight line?
Whereas a single-line segment is sufficient to make a straight-line hypothesis, the authors also use a pair of line segment for robust fitting.
Q16. Why is the second control point examined?
The second control point is also examined to see if its position is too low, because if it keeps going down, it will eventually collide with the first control point.
Q17. What is the way to detect a curb?
Whether curbs can be detected or not depends on the application—detecting a single lane boundary is sufficient in many applications, including the ones for collision warning.