«上一篇
 文章快速检索 高级检索

 应用科技  2020, Vol. 47 Issue (2): 35-43  DOI: 10.11991/yykj.201906016 0

### 引用本文

HU Qiang, QU Qiang, HE Xin. An improved sidewalk detection algorithm based on multi-feature fusion[J]. Applied Science and Technology, 2020, 47(2): 35-43. DOI: 10.11991/yykj.201906016.

### 文章历史

1. 南京航空航天大学 自动化学院，江苏 南京 211106;
2. 南京航空航天大学 计算机科学与技术学院，江苏 南京 211106

An improved sidewalk detection algorithm based on multi-feature fusion
HU Qiang1, QU Qiang1, HE Xin2
1. College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China;
2. College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China
Abstract: Research on road detection is mostly based on open datasets of lane such as KITTI. Due to the differences in color, material and surrounding environment between lanes and sidewalks, accurate detection of the sidewalk area is also a problem that needs to be solved. Therefore, the application scenario is set as an outdoor sidewalk in this paper, proposing an improved sidewalk detection algorithm based on multi-feature fusion. Firstly, SLIC super-pixel algorithm is used to obtain super-pixel image to reduce the noise interference and follow-up training dimension. Then, calculating each super-pixel block feature, a new texture extraction method based on Gabor filter is proposed to reduce the time complexity, and the principal components analysis (PCA)-based illumination invariant spatial features and three-dimensional depth gradient features are added to improve detection accuracy. The Adaboost classifier is used to train the fused feature vectors and predict the sidewalk area. Finally, the Markov random field is used to optimize the segmentation results. This method is universal and does not depend on a priori road appearance and structure. The validity of the algorithm is verified by the experiment based on the created sidewalk datasets.
Keywords: sidewalk detection    super-pixel    Gabor texture    illumination invariant space    three-dimensional depth    multi-feature fusion    machine learning    Markov random field

1 多特征融合的人行道检测算法

1.1 多特征融合

 ${v_{{\rm{fusion}}}} = \left( {{v_{{\rm{2D}}}},{v_{{\rm{3D}}}}} \right)$

 $\left\{ \begin{array}{l}\!\!{u_i} = {u_0} + \dfrac{{fX - {\varepsilon _i}\dfrac{b}{2}f}}{{(Y + h)\sin \theta + Z\cos \theta }}\\ \!\!v = {v_0} - f\tan \theta + \dfrac{{f(Y + h)}}{{\cos \theta [(Y + h)\sin \theta + Z\cos \theta ]}}\end{array} \right.$ (1)

 $\varDelta = \frac{{fb}}{{(Y + h)\sin \theta + Z\cos \theta }}$ (2)

 $Y = \frac{{b(v\cos \theta - {v_0}\cos \theta + f\sin \theta )}}{\varDelta } - h$

1.2 改进的Gabor滤波器纹理提取算法

 ${g_{\phi ,{\omega _0}}}(x,y) = \frac{{{\omega _0}}}{{\sqrt {2{{\text{π}}}} c}}{{\rm{e}}^{ - {\omega _0}^2(4{a^2} + {b^2})/8{c^2}}}({{\rm{e}}^{ia{w_0}}} - {{\rm{e}}^{ - {c^2}/2}})$

 ${I_{\phi ,{\omega _0}}}(p) = I(p) \otimes {g_{\phi ,{\omega _0}}}(p)$

 ${E_{\phi ,{\omega _0}}}(p) = \sqrt {\operatorname{Re} {{\left( {{I_{\phi ,{\omega _0}}}\left( p \right)} \right)}^2} + \operatorname{Im} {{\left( {{I_{\phi ,{\omega _0}}}\left( p \right)} \right)}^2}}$ (3)

 $\begin{split} & {E^1}_{\phi ,{\omega _0}}\left( k \right) > {E^2}_{\phi ,{\omega _0}}\left( k \right) > {E^3}_{\phi ,{\omega _0}}\left( k \right) > {E^4}_{\phi ,{\omega _0}}\left( k \right) > \\ & {E^5}_{\phi ,{\omega _0}}\left( k \right) > {E^6}_{\phi ,{\omega _0}}\left( k \right) > {E^7}_{\phi ,{\omega _0}}\left( k \right) > {E^8}_{\phi ,{\omega _0}}\left( k \right) \end{split}$

 ${\rm{Conf}}\left( k \right) = \left\{ \begin{array}{l} \!\!{1 - \dfrac{{{E^8}_{\phi ,{\omega _0}}\left( k \right)}}{{{E^1}_{\phi ,{\omega _0}}\left( k \right)}},\quad {E^1}_{\phi ,{\omega _0}}\left( k \right) > {E_{{\rm{th}}}}}\\ \!\!{0,\quad {E^1}_{\phi ,{\omega _0}}\left( k \right) \leqslant {E_{{\rm{th}}}}} \end{array} \right.$

1.3 基于PCA的光照不变空间算法

 $\left\{ \begin{array}{l} \!\!r = \log \left( {\dfrac{R}{G}} \right) \\ \!\!b = \log \left( {\dfrac{B}{G}} \right) \end{array} \right.$

 $\left\{ \begin{array}{l} \!\!r = \log \left( {\dfrac{R}{{{{\left( {RGB} \right)}^{1/3}}}}} \right) \\ \!\!b = \log \left( {\dfrac{B}{{{{\left( {RGB} \right)}^{1/3}}}}} \right) \end{array} \right.$

 ${{C}} = {{X}}{{{X}}^{\rm{T}}} = \left[ {\begin{array}{*{20}{c}} {{{{e}}_1}}&{{{{e}}_2}} \end{array}} \right]\left[ {\begin{array}{*{20}{c}} {{\lambda _1}}&0 \\ 0&{{\lambda _2}} \end{array}} \right]\left[ {\begin{array}{*{20}{c}} {{{{e}}_1}^{\rm{T}}} \\ {{{{e}}_2}^{\rm{T}}} \end{array}} \right]$

2 实验分析与验证 2.1 数据收集与标注

2.2 超像素算法评估

SLIC具有唯一参数超像素块数K，为了验证该参数对人行道检测精度的影响，该实验基于RGB+POS+HSV特征组合，K从100~1 000间隔为10地选取数值，对每张图像训练时间和超像素块检测精度进行统计，如图8所示。

2.3 改进的Gabor纹理提取算法加速效果评估

2.4 多特征融合评估

RGB和HSV表示的是颜色信息，RGB比HSV识别道路的能力更好，Gabor提供的是纹理信息，POS表达的是位置信息，GKDES的检测结果最差，PCA-II和DGKDES是新加入的特征，因此首先选择4种组合方式：“RGB+Gabor”、“HSV+POS”、“HSV+Gabor”和“RGB+Gabor+POS”。根据表3的评估结果说明“RGB+Gabor+POS”组合的表现最好。

2.5 MRF优化结果和不同算法对比实验