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

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.