基于视觉显著性的煤矿井下关键目标对象实时感知研究

Study on real-time perception of target ROI in underground coal mines based on visual saliency

随着煤矿智能化技术发展,井下关键设备目标对象视觉感知应用需求日益增强。井下复杂场景,特别是生产工况综采工作面,人员及设备目标频繁交错呈现。基于监控视觉画面实时检测、提取人员及关键设备目标对象,对实现井下关键设备目标对象智能监控,生产场景智能感知与安全生产管理意义重大,因此需要研究井下关键目标对象实时感知方法。基于视觉注意机制的显著目标检测和分割是复杂场景关键目标对象感知的有效方法之一,但是显著性检测和目标分割过程计算复杂度高、耗时长,难以达到工程应用的实时性要求。基于此,在分析图像视觉特征的基础上,特别是煤矿井下图像视觉特征,提出一种基于随机采样区域对比度计算的实时显著性检测方法,引入随机采样策略对原图像像素进行采样后利用Efficient Graph-based Segmentation方法将图像分割为若干区域,然后计算区域对比度获得区域显著性,实现了实时显著性检测;在显著性区域或者目标分割过程中,提出一种自适应的前景背景阈值迭代方法,基于Shared Sample Matting方法实现显著目标的实时分割提取。基于公共数据集进行试验分析,结果表明,该方法不仅提高了显著性目标的检测分割精度,而且达到30 FPS左右的显著目标检测、分割实时处理效率。同时,将该方法应用于煤矿井下复杂场景中关键设备目标对象的实时感知,效果良好,满足工程应用需求。

With the development of intelligent technology in underground coal mines, the demand of visual perception applications for target objects of key equipment is increasing. In complex underground scenes, especially in the fully-mechanized mining face of producion conditions, where the target objects of personnel and equipment are frequently staggered. Based on the visual monitoring image, real-time detection and extraction of the targets objects is critical to achieve the intelligent monitoring of the key objects, the intelligent perception of production scenarios and safety production management in underground coal mine. Therefore, it is of great significance to study the method of the real-time perception of the key objects in underground coal mine. Salient object detection and segmentation based on visual attention mechanism is one of the effective way to target ROI perception in complex scene. However, the salient detection and object segmentation processes are computationally complex and time-consuming, making it difficult to reach the real-time requirements of engineering applications. Therefore, on the basis of analysis of image visual features, especially the image visual features of underground coal mines, a real-time saliency detection method based on the random sampling region contrast was proposed. The random sampling strategy is introduced to sample the original image pixels and then Efficient Graph is applied to segment the image into regions. Then region contrast is calculated as saliency to achieve real-time salient detection. In the process of salient object segmentation, an adaptive foreground and background threshold iteration method is proposed for the real-time salient object segmentation by the Shared Sample Matting method. Quantitative comparisons with state-of-the-art methods on benchmark datasets are carried out, and the experimental analysis based on the public data set shows that the method not only improves the accuracy of saliency detection and salientobject segmentation, but also achieves the real-time processing efficiency of salient object detection and segmentation at about 30 FPS. At the same time, the proposed method is applied to the real-time perception of the target objects in the underground coal mine with high performance, and meets the industrial application requirements.