Densely Stacked Rebar Counting Based on Image Semantic Segmentation in Wide Scene

To accurately count densely stacked rebars belongs to semantic segmentation in image. Various methods have been proposed to segment the objects in single image. However, widely and densely stacking scene such as warehouse or vehicle stacking of rebars is always faced. For these scenes, clear image of end faces of rebars cannot be obtained through a single camera. Generally, multiple cameras are often used to take clear images to count rebars. One strategy is to stitch multiple images into a single image at pixel level first, and then count rebars in this stitched image. Various methods have been proposed to stitch images, but it is extremely difficult to stitch two images at pixel level accurately for parallax distortion or partially hidden. A topological stitching strategy is proposed in this paper to count the rebars in whole scene. In the method, rebars in each image are identified and counted first, and then the identification and counting results in the adjacent images are stitched so as to esitmate the total number of rebars. The perfomance of the method is evaluated on real datasets, and the counting accuracy exceeds 99%. Experiments show that the proposed method is effective on widely and densely stacked rebars identification.