Pyr8 travels CV world

[Review] Real-Time Lane-Vehicle Detection and Tracking System

Lane Detection Tracking Article Review
Word count: 407Reading time: 2 min
2019/10/06 Share

framework

  1. lane detection and tracking
  2. sky and road region recognition
  3. vehicle detection and tracking

intro

This paper is inspired by the advantages of object detector based on CNN, such as Faster-RCNN, YOLO, and so on. Therefore, car detection is the core step in the framework, whereas the lane detection and tracking is used to reduce ROI regions.

1. Lane detection and tracking

This traditional pipeline is adopted:

    graph LR
    subgraph detect
    1[RGB] --> 2[Gray]
    2 --> 3[Gauss Blur]
    3 --> 4[Canny]
    4 --> 5[Hough Transform]
    end
    5 -.-> 6
    subgraph split
    6[Slope filter] --> 7[left/right split]
    end
    7 -.-> 8
    subgraph track
    8[change compute] --> 9[update lanes]
    end

2. Sky and road region recognition

Road region: 

    graph LR
    a[lane information] --> b[Canny edge]
    b --> c[region growing from bottom]

Sky region: 

    graph LR
    1[vertical mean distribution] --> 2[threshold]

3. Vehicle detection

A simpler CNN to be more easily converged is proposed, which does not predict the coordinates of the object bounding boxes directly. Otherwise, the b-boxes are K centroids from clustering of ground truth, and DPM regression is used to adjust them.

Moreover, with box confidences, the loss is weighted sum of the regression and classification.

Then, the authors declared three innovations:

  1. A general preprocessing is adopted.
  2. The loss of the regression part is improved, by assigning any boxes with a larger overlap than the threshold to the ground truth.
  3. Negative samples are added.

CNN Model: Alex-Net with replaced top classification layer.

Training process: 

    graph LR
    1[300 videos on highway] --> 2[5000/1000 vehicle images for train/test]
  1. multi-scale
  2. overlap < 0.2 as negative | overlap > 0.8 as positive
  3. Aspect ratios in range of [0.8, 1.2] are accepted, and paddings are performed
  4. 30000 epochs, batch size of 128, momentum of 0.9 weight decay 0.0005, lr 0.001/20000 and 0.0001/10000

4. Vehicle tracking

Improved FB-error tracking framework based on sparse Pyramid optical flow is used2.

5. Result

65% AP and 90% Recall, nothing impressive.

Reference

1. Guan, Huang , et al. “Real-Time Lane-Vehicle Detection and Tracking System,” 2016 Chinese Control and Decision Conference (CCDC) IEEE, 2016. ↩
2. P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan, “Object detection with discriminatively trained part-based models,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 32, pp. 1627-1645, 2010. ↩

Author：Henry Yang

原文链接：http://guigua1.github.io/2019/10/Lane-Vehicle-Detection-Tracking/

发表日期：October 6th 2019, 12:00:00 am

更新日期：January 9th 2020, 7:49:16 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. Lane detection and tracking
  4. 4. 2. Sky and road region recognition
  5. 5. 3. Vehicle detection
  6. 6. 4. Vehicle tracking
  7. 7. 5. Result
    1. 7.1. Reference
Total : 8
2019
Lane Detection Tracking Article Review Point Cloud Segmentation Automated Driving Vanishing Point CNN
Lane-Detection Point-Cloud