Real-time evaluation of the blending uniformity of industrially produced gravelly soil based on Cond-YOLOv8-seg

Abstract

Industrial production of gravelly soil has been applied at construction sites in earth-rock engineering, which adopts automatic blending and belt-conveyer systems for gravelly soil. The blending uniformity of gravelly soil has a significant impact on its mechanical and physical properties including shear resistance and impermeability. Real-time evaluation of blending uniformity is key to controlling the quality of gravelly soil. This study proposes a real-time evaluation method for the blending uniformity of gravelly soil transported by a belt conveyor. First, Cond-YOLOv8-seg is adapted to accurately segment gravel particles from gravelly soil images, employing a conditional convolution segmentation head and an improved Protonet embedded with a recurrent criss-cross attention (RCCA) module based on YOLOv8-seg. Then, the segmented masks are used to analyze the spatial uniformity (SU) and blending ratio consistency (BRC) of individual and continuous gravelly soil images, employing the proposed evaluation criterion for SU based on the theory of deformable four-sided static distances and BRC based on the area ratio of different-scale gravel particles in one image. Cond-YOLOv8-seg achieved an AP50 of 0.803 and FPS of 58.3. The effectiveness of the proposed evaluation criteria is validated quantitatively. The average uniformity evaluation time of a 2048 × 1536 image is determined to be 0.1151s, which is close to the real-time evaluation of gravelly soil transported by a belt conveyor.