Autonomous soil vision scanning system for intelligent subgrade compaction

The wide application of intelligent compaction in subgrade construction is limited by the evaluation accuracy. The current practice assumes the compacted site as a homogeneous soil condition, ignoring the influence of soil property. This paper describes an advanced autonomous soil scanning system based on the computer-vision approach to recognize the soil gradation in real-time. A hybrid intelligent model is developed through a series of field compaction tests, integrating Deeplabv3+ and XGBoost algorithms. The proposed method is effective in segmenting and quantifying the large particles while the fine-grained soil is completed by the prediction model. The scanned soil information correlates well with the intelligent compaction indicators, providing essential parameters for the accurate assessment of subgrade compaction quality. The soil vision scanning system is combined with the intelligent compaction system to enable the roller autonomous learning of the whole con-struction area.