Tracking of real aggregates in asphalt mixture using a modified convolutional neural network and multi-feature similarity weighted matching

The macroscopic performance of asphalt mixtures depends on aggregate movement and rearrangement during compaction, while global tracking of this motion remains challenging. This study develops a novel method for tracking real aggregate migration using CT scanning and 3D reconstruction. A convolutional neural network integrating a Residual network, a Squeeze-Excitation module, and a Nested U-Net architecture (RSNU) was constructed for precise aggregate segmentation under various compaction states. Subsequently, a tracking algorithm employing K-Dimensional Tree (KD-Tree) search and Multi-feature similarity Weighted Matching (KMWM) was developed, with weights optimized via grid search. Validation using a compaction sequence that included preset rigid transformations and morphological errors demonstrated that: The RSNU achieved highly consistent segmentation across gradations and compaction states. The KMWM demonstrated strong robustness, with a chain tracking match rate greater than 95% and only 0.29% of displacement errors exceeding the voxel size. Under actual compaction conditions, the global chain tracking rates for Asphalt Concrete (AC), Stone Mastic Asphalt (SMA), and Open-Graded Friction Course (OGFC) reached 93.20%, 92.91%, and 91.36%, respectively. This study provides a novel approach for quantitatively studying the compaction mechanism of asphalt pavement at the meso-structural level.