Abstract: Existing coal flow calculation methods generally treat coal bulk density as a constant, resulting in inaccurate results. This paper employs a depth camera to collect and fuse RGB and point cloud data, enabling precise measurement of dynamic coal flow. First, coal flows are classified into five density grades according to national standards, with experimental determination of density values corresponding to different coal particle shapes at each grade. To enhance computational efficiency, a small-sample dataset of coal flows composed of particles of varying sizes is constructed and labeled with multiple classes. Subsequently, an optimized LW-UNet3+ (Lightweight-UNet3+) segmentation algorithm is proposed. This lightweight architecture incorporates multi-scale attention and adaptive multi-receptive field modules to perform multi-class segmentation of coal flow images. Based on this segmentation, the surface pixel proportion and overall density level of coal particles in each density grade are calculated. Subsequently, a pixel-stitching method based on unit coal flows was developed to obtain unit coal flow data and corresponding point cloud data. An optimized Delaunay triangulation algorithm was employed to calculate coal flow volume, from which the coal flow rate was derived. Cumulative summation of unit coal flow rates yielded the total coal flow rate. Experimental results demonstrate that the improved method achieves over 15% higher accuracy compared to traditional approaches.