Geometric-aware instance segmentation for robust industrial monitoring of conveyor belt deviation

Belt conveyor misalignment is a common fault in mining and mineral processing, threatening equipment safety and production efficiency. Traditional vision-based detection methods suffer from unstable references and poor performance in complex industrial environments. This work proposes a geometric-constrained instance segmentation framework combined with dedicated physical markers for real-time embedded misalignment monitoring. A distance-aware loss and a boundary-enhanced module are designed to improve pixel-domain geometric consistency, specifically by refining arrow-tip localization, reducing output fluctuation, and enhancing the linear response under synthetic displacement injection. The system is deployed on an RK3588 edge platform with INT8 quantization, achieving an end-to-end throughput of approximately 26 FPS and a coefficient of determination of 0.9862 in synthetic-displacement linearity tests. The method presents a promising visual-computing approach for industrial conveyor monitoring under harsh conditions; however, the current implementation has been validated only under the evaluated conditions, and physical (millimetre-level) calibration and an independent test set warrant further validation. The source code and trained models are available at https://github.com/29x9vqp7sk-creator/ConveyorBelt_Monitoring and permanently archived at Zenodo with https://doi.org/10.5281/zenodo.22138817 .