A Large-Scale Cross-View Truck Benchmark with Dynamic Adaptive Fusion Attention Mechanism

Cross-View Truck Re-Identification (CV-TR) constitutes a pivotal component in intelligent urban traffic systems, particularly under non-cooperative surveillance scenarios where license plate information or aerial imagery becomes inaccessible. This specialized visual intelligence task exhibits two fundamental technical barriers that diverge fundamentally from general object re-identification (Re-ID): 1) Fine-grained discrimination, which required between visually similar truck subtypes, and 2) Single-view, which limitations in capturing multi-angle features critical for robust cross-camera matching. To bridge this gap, we propose two key contributions: First, we introduce a CV-TR dataset, a large-scale benchmark comprising 62, 783 annotated truck images collected from three real-world traffic surveillance scenarios. It includes 2, 683 unique truck identities, each documented from front, side, and rear viewpoints. Second, we develop the Dynamic Adaptive Cross-View Network (DACVNet), which features a vehicle-centric architecture with adaptive feature fusion attention mechanism to tackle large-scale cross-view matching challenges. Experiments show DACVNet achieves state-of-the-art performance, significantly surpassing existing approaches in cross-view truck Re-ID. The CV-TR dataset is available at https: //pan. baidu. com/s/11LF0vIB7hCJQ9ᵣdpNNs2A? pwd=eqt7.