Vision-based truck unloading-state recognition in large dam scenes

Efficient monitoring of truck unloading on dam surfaces is challenged by substantial target-scale variation, partial occlusion and the need to maintain identity-consistent trajectories in large construction scenes. This study proposes a tracking-guided fine-grained recognition framework integrating YOLOv8–DeepSORT multi-object tracking with a self-boosting attention mechanism (SAM) enhanced by an asymmetric dual-branch channel-attention module. YOLOv8 detects haul trucks, while DeepSORT combines motion and appearance features to maintain identity-consistent trajectories and bridge short-term detection failures caused by distance or occlusion. The two channel branches use different reduction ratios and are integrated with SAM to strengthen discriminative unloading-state feature. Experiments on real dam-surface monitoring videos show that the proposed module improves unloading classification accuracy from 0.932 to 0.975. In field deployment, the framework achieves an average Multiple Object Tracking Accuracy (MOTA) of 0.998 and an average unloading recognition accuracy of 0.9217. The extracted trajectories and unloading states further support quantitative estimation of transportation and unloading durations, providing data for construction-efficiency assessment and operational planning.