Pyr8 travels CV world

[Review] Multiple Lane Boundary Detection Using a Combination of Low-Level Im...

Lane Detection Article Review
Word count: 463Reading time: 2 min
2019/10/01 Share

framework

  1. Fusion of edge features and high intensity feature
  2. Application of Kalman filter for information fusion and tracking
  3. Test on roads with different driving scenarios, including day, night, heavy traffic, rain, confusing textures and shadows

intro

Detection multiple lanes and achieving reliability under different environments are remaining challenges. At first, for vision-based feature extraction, intensity is unreliable and should be combined with other constraints, such as parallel boundaries and intersect around vanishing point. However, without a priori, performance often degrades.

Therefore, edge feature and high intensity feature are fused to enhance the stability of feature measurement. Variance of measurement noise is estimated live and indicates the corresponding interference. Moreover, Kalman filter is used in case that both measurements are contaminated.

1. Edge Feature Extraction

graph LR
subgraph Edge
edge --fitting--> line
line --grouping--> boundary
end
subgraph Intensity
boundary --searching--> 1[high intensity]
1 --refining--> lane
end

  1. ROI is defined below a hand-calibrated horizontal line.

  2. Canny + Hough Transform

  3. Support lines $y^1$ and $y^2$ below the horizontal line, and intersections with edge lines

  4. Edge feature: $\mathbf{I}_i^e = [ x_i^1, x_i^2]^T \in \mathbb{R}^2$

  5. Grouping:

    1. Sort $\mathbf{I}_i^e$ by slope $1 / (x_i^1 - x_i^2)$
    2. Compute distance $\Delta\mathbf{d}{i,i+1} = [x_i^1-\{i+1}^2|, |xi^2-x{i+1}^2|]$
    3. Compare $\Delta\mathbf{d}_{ij}$ with threshold $t=[\frac{w}{2n_1}, \frac{w}{2n_2}]$, where $w$ is the width of the image frame, and $n_1, n_2$ is the maximum number of lanes that can be held in $w$ on $y^1, y^2$, respectively.
    4. Average lines: $\hat{\mathbf{I}}k^e = \frac{1}{n_k}\sum{l=1}^{nk}\mathbf{I}{i+l}^e$

  6. Searching local maxima features at intervals along boundary

    $d$ horizontal pixels beside the boundary are convolved with a first derivative Gaussian Kernel

    also, $\sigma$ varies from 31 for the lower half boundary to 21 for the upper half.

  7. Smoothness score of local maxima feature

  8. RANSAC refining with $\gamma_k$ lower than a certain value, get $\mathbf{I}_i^m$

2. Kalman Filter

state vector: lane intersections on the two support lines, lateral speed of boundary movement on the support lines, and interference term.

3. Performance

data: 1024x480 pixels at 30 fps

Avg. Precision: 90.9% Avg. Recall: 91.0%

Processing Time: 0.0130 sec per frame

5. Review

Many priori parameters are implicit in this paper, so that it is hard to recall consistent performance for my experiments.

Reference

Author：Henry Yang

原文链接：http://guigua1.github.io/2019/10/multi-lane-bndry/

发表日期：October 1st 2019, 12:00:00 am

更新日期：January 9th 2020, 8:00:41 pm

版权声明：本文采用知识共享署名-非商业性使用 4.0 国际许可协议进行许可

CATALOG
  1. 1. framework
  2. 2. intro
  3. 3. 1. Edge Feature Extraction
  4. 4. 2. Kalman Filter
  5. 5. 3. Performance
  6. 6. 5. Review
  7. 7. Reference
Total : 8
2019
Lane Detection Tracking Article Review Point Cloud Segmentation Automated Driving Vanishing Point CNN
Lane-Detection Point-Cloud