﻿ 基于多特征融合和条件随机场的道路分割
 计算机系统应用  2020, Vol. 29 Issue (3): 240-245 PDF

Road Segmentation Method Based on Multi-Feature Fusion and Conditional Random Field
YAN Zhao-Fan, LI Yu-Chong, YAN Guo-Ping
School of Information Engineering, Chang’an University, Xi’an 710064, China
Foundation item: “Hongyi Prospers” Graduates Innovative Research Practice Project (2018103, 2018109)
Abstract: In the complex traffic scene image, road segmentation is difficult and the edges of the segmentation are rough. In order to solve this problem, a road segmentation method based on multi-feature fusion and conditional random field is proposed. Firstly, the textons and color features of the image are extracted from the traffic image. Then, the road segmentation problem is regarded as a pixel-based binary classification problem. The extracted texton features and color features are fused and input into the SVM classifier, which can achieve the coarse segmentation of the road area and the background area in the traffic image. Finally, by using the color and position constraints of the fully connected conditional random field to optimize segmentation results, a smoother segmentation edge can be obtained and compared with other segmentation algorithms. The experimental results demonstrate that road segmentation method that based on the multi-feature fusion and the conditional random field achieves 95.37% of average segmentation accuracy and 94.55% of mean pixel accuracy.
Key words: image pattern recognition     road segmentation     texton     multi-feature fusion     conditional random field

1 基于多特征融合的条件随机场

 图 1 基于多特征融合与条件随机场的道路分割模型

1.1 纹理基元特征

 图 2 条件随机场与分割结果的关系图

 图 3 交通场景图像的纹理基元图

17维的纹理基元滤波器组是由不同尺度的高斯平滑滤波器, 高斯差分滤波器, 高斯拉普拉斯滤波器组成的. 3种滤波器在尺度为k时分别被定义为:

 $G(u,v) = \frac{1}{{2\pi {k^2}}}\exp \left( - \frac{{{u^2} + {v^2}}}{{2{k^2}}}\right)$ (1)
 $LOG(u,v) = \frac{1}{{\pi {k^4}}}\left(\frac{{{u^2} + {v^2}}}{{2{k^2}}} - 1\right)\exp \left( - \frac{{{u^2} + {v^2}}}{{2{k^2}}}\right)$ (2)
 ${G_x}(u,v) = \frac{\partial }{{\partial u}}G(u,v)$ (3)

 图 4 17维滤波器组

1.2 颜色特征与特征融合

 $F = [{F_t},F_{{c}}^{\rm RGB}]$ (4)

1.3 条件随机场

 $P(X|I) = \frac{1}{{Z(I)}}\exp ( - E(X|I))$ (5)

 $E(X) = \sum\limits_{i} {{\psi _u}({x_i})} + \sum\limits_{i < j} {{\psi _p}({x_i},{x_j})}$ (6)

 ${\psi _u}({x_i}) = - \log p({x_i})$ (7)

 \begin{aligned}[b] {\psi _{^p}}({x_i},{x_j}) =& \mu ({x_i},{x_j})\left[ {\omega ^{(1)}}\exp \left( - \frac{{|{p_i} - {p_j}|2}}{{{\theta _\alpha }}} - \right. \frac{{|{I_i} - {I_j}{|^2}}}{{{\theta _\beta }}}\right) \\ &\left. +{\omega ^{(2)}}\exp \left( - \frac{{|{p_i} - {p_j}{|^2}}}{{{\theta _\gamma }}}\right)\right] \end{aligned}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! (8)

2 实验与分析 2.1 实验数据

2.2 评价指标

 $MPA = \frac{1}{N}\sum\limits_{i} {\frac{{{N_{ai}}}}{{{N_{ti}}}}}$ (9)

 $TPR = \frac{{TP}}{{TP + FP}}$ (10)
 $FPR = \frac{{FP}}{{TN + FP}}$ (11)

TPRFPR的取值范围为0到1,TPR的值越大, FPR的值越小表示模型的性能越好. 对应于AUC值的取值范围为0.5到1, 表示模型的平均分割准确率. 当取值为1时, 模型的性能最好, 表示所有的像素全部被正确的分类.

2.3 实验结果

 图 5 本文使用方法与仅使用一元势能方法的分割结果比较

3 总结

 [1] Costea AD, Nedevschi S. Fast traffic scene segmentation using multi-range features from multi-resolution filtered and spatial context channels. 2016 IEEE Intelligent Vehicles Symposium. Gothenburg, Sweden. 2016. 328–334. [2] Hillel AB, Lerner R, Levi D, et al. Recent progress in road and lane detection: A survey. Machine Vision and Applications, 2014, 25(3): 727-745. DOI:10.1007/s00138-011-0404-2 [3] Li Y, Xu LL, Rao J, et al. A Y-Net deep learning method for road segmentation using high-resolution visible remote sensing images. Remote Sensing Letters, 2019, 10(4): 381-390. DOI:10.1080/2150704X.2018.1557791 [4] 田峥, 徐成, 米超, 等. 基于消失点和主方向估计的道路分割算法. 计算机研究与发展, 2014, 51(4): 762-772. [5] 吕书强, 张思博. 基于阈值分割和形态学的高分辨率遥感影像道路提取. 北京建筑工程学院学报, 2012, 28(1): 35-39. [6] 吴实, 熊淑华, 李杰, 等. 基于简单线性迭代聚类算法和动态区域合并的岩屑图像分割. 科学技术与工程, 2016, 16(35): 238-243. [7] Zhou HL, Kong H, Wei L, et al. Efficient road detection and tracking for unmanned aerial vehicle. IEEE Transactions on Intelligent Transportation Systems, 2015, 16(1): 297-309. DOI:10.1109/TITS.2014.2331353 [8] Alvarez JM, Gevers T, Lopez AM. Road detection by one-class color classification: Dataset and experiments. arXiv: 1412.3506, 2014. [9] Chen LC, Papandreou G, Kokkinos I, et al. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 40(4): 834-848. [10] 尹艳鹏, 周颖, 曾丹, 等. 基于多特征融合条件随机场的人脸图像分割. 电子测量技术, 2015, 38(6): 54-59. [11] He XM, Zemel RS, Carreira-Perpinan MA. Multiscale conditional random fields for image labeling. Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. Washington, DC, USA. 2004. [12] Krähenbühl P, Koltun V. Efficient inference in fully connected CRFs with Gaussian edge potentials. arXiv: 1210.5644, 2012. [13] Shelhamer E, Long J, Darrell T. Fully convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(4): 640-651. [14] Chen LC, Papandreou G, Kokkinos I. Semantic image segmentation with deep convolutional nets and fully connected CRFs. Compute Science, 2014, 4: 357-361. [15] Payet N, Todorovic S. (RF)2-random forest random field. Advances in Neural Information Processing Systems (NIPS). 2010. [16] Shotton J, Winn J, Rother C, et al. TextonBoost for image understanding: multi-class object recognition and segmentation by jointly modeling texture, layout, and context. International Journal of Computer Vision, 2009, 81(1): 2-23. DOI:10.1007/s11263-007-0109-1 [17] Winn JM, Criminisi A, Minka TP. Object categorization by learned universal visual dictionary. 10th IEEE International Conference on Computer Vision. Beijing, China. 2005. Vol 2. 1800–1807. [18] Zhang LG, Verma B. Roadside vegetation segmentation with adaptive texton clustering model. Engineering Applications of Artificial Intelligence, 2019, 77: 159-176. DOI:10.1016/j.engappai.2018.10.009 [19] Wu J, Feng L, Liu SL, et al. Image retrieval framework based on texton uniform descriptor and modified manifold ranking. Journal of Visual Communication and Image Representation, 2017, 49: 78-88. DOI:10.1016/j.jvcir.2017.08.002 [20] 宋青松, 张超, 陈禹, 等. 组合全卷积神经网络和条件随机场的道路分割. 清华大学学报(自然科学版), 2018, 58(8): 725-731.