Real-Time Edge-Prior Guided SegFormer for Robust Contour Extraction of Aggregate Particles in Conveyor-Belt Depth Maps

Accurate contour extraction of aggregate particles from conveyor-belt depth maps is essential for downstream particle counting and size measurement, yet industrial depth data often contains weak discontinuities, missing values, and speckle-like noise. We propose a task-specific geometry-aware contour extraction framework that combines a compact SegFormer encoder with depth-derived priors, a lightweight local branch, edge-prior gated fusion, and full-resolution residual refinement. The input representation consists of normalized depth, Sobel gradient magnitude, and the absolute Laplacian response. On AGG_FULLDATA, the method achieves Optimal Dataset Scale (ODS), Optimal Image Scale (OIS), and Average Precision (AP) values of 0.9607/0.9716/0.9683 under the primary tolerance-based protocol (tol=1), while retaining an ODS of 0.6476 under strict pixel-exact matching. On External130, a test-only split collected under altered operating conditions using the same sensor, it reaches 0.9580/0.9734/0.9683 without retraining and consistently outperforms the MiT-only baseline. A rigid-object repeatability study based on 30 raw PLY scans shows a mean boundary deviation of 0.335 px, a within-1 px correspondence rate of 97.1%, and a coefficient of variation (CV) of equivalent diameter below 1%, supporting the practical meaning of tol=1. The full pipeline runs at 48.9 frames per second (FPS) with 3.71M parameters on an NVIDIA GeForce RTX 4060 GPU. Broader robustness to separately controlled operating factors, environmental disturbances, and cross-device settings still requires validation.