安全帽分割是实现煤矿人员智能视频监控的关键技术之一，可促进人员定位、跟踪、安全帽佩戴检测等相关技术的研究，为此，提出一种基于超像素特征提取与支持向量机(Support Vector Machines,SVM)分类的矿井人员安全帽分割方法。首先，采用简单线性迭代聚类(Simple Linear Iterative Clustering,SLIC)模型将人员图像粒化为一定数量内部像素点颜色特征相似且空间位置相近的超像素。其次，提取超像素在RGB，YCbCr，Lab，HSV空间上的颜色特征及其灰度直方图纹理特征，通过分析安全帽二维轮廓线上的斜率变化特性建立安全帽轮廓特征模型。最后，在训练集人员图像中分别提取安全帽正样本超像素和背景负样本超像素的颜色、纹理特征训练SVM分类器，采用已训练的SVM将测试集中的人员图像超像素二分类为安全帽正样本和背景负样本。进一步通过安全帽轮廓特征模型判别SVM误分类的虚假正样本并对其进行类别修正，识别同时包含正样本像素点和负样本像素点的欠分割样本超像素，并通过求取正样本区域边界掩模与Prewitt算子所提取轮廓的差集对其进行二级像素分类，分离出其中的正样本像素区域，得到安全帽的最终分割结果。文中对测试数据集中的100幅人员图像进行实验，结果表明:所提安全帽分割算法的精确度Pr(Precision)均值达到96.94%，召回率Re(Recall)均值达到95.83%，对不同场景下的人员安全帽图像具有较好的适用性与鲁棒性。

Safety helmet segmentation is one of the key technologies to realize intelligent video surveillance of personnel in coal mine,which can promote the research of personnel positioning,tracking,and safety helmet wearing detection and other related technologies.For this reason,a method of mine personnel safety helmet segmentation based on super pixel feature extraction and support vector machines (SVM) classification is proposed.Firstly,the simple linear iterative clustering (SLIC) model is used to granulate the personnel image into a certain number of super pixels with similar internal pixel color features and similar spatial locations.Secondly,the color features and gray histogram texture features of super pixels in RGB,YcbCr,Lab and HSV space are extracted,and the profile feature model of safety helmet is established by analyzing the slope variation characteristics of the two-dimensional contour line of the safety helmet.Finally,the color and texture features of the positive sample super pixels of the helmet and the background negative sample super pixels are extracted from the personnel images of training set to train the SVM classifier,and the trained SVM is used to classify the super pixels of the personnel images in the testing set into the positive samples of helmet and negative samples of background.Furthermore,the fake positive samples misclassified by the SVM are identified by the helmet profile feature model and the category correction is performed for them,the under segmented sample super pixels contain both positive sample pixels and negative sample pixels are recognized,they are processed as secondary pixel classification by obtaining the difference set between the boundary mask of the positive sample region and the contour extracted by the Prewitt operator,and the helmet pixels region of the positive sample is separated to obtain the final segmentation result of the helmet.The experiments are carried out on 100 personnel images in the testing data set,and the results show that the average precision of the proposed helmet segmentation algorithm reaches 96.94%,the average recall rate reaches 95.83%,the algorithm has good applicability and robustness for personnel helmet images in different scenarios.

1 人员图像超像素粒化

2 超像素特征提取

   2.1 颜色特征

   2.2 纹理特征

   2.3 轮廓特征

3 SVM分类超像素

   3.1 SVM分类器训练及预测分类

   3.2 超像素类别修正

   3.3 欠分割样本超像素处理

4 实验结果分析

5 结论