Low‐complexity real‐time detection Transformer for identifying bridge vehicle loads

Vehicle load identification (VLI) is pivotal for bridge health monitoring, safety assessment, and intelligent maintenance. However, computer vision-based VLI is confronted by two critical challenges, that is, compromised identification accuracy under dynamic scene and computational constraints imposed by edge monitoring devices. To this end, a low-complexity real-time detection Transformer (LC-RTDETR) is developed to establish a framework for bridge VLI. The proposed LC-RTDETR provides foundational perception for VLI and features three advantages: (1) lightweight feature extraction via the star network backbone, (2) robust feature representation enabled by the dynamic-range histogram self-attention module for single-scale fusion, and (3) enhanced multi-scale processing efficiency through the proposed context-guided spatial feature reconstruction pyramid network. This architecture augments accuracy in complex scenes while reducing computational demands. For continuous trajectory acquisition, detections from the proposed LC-RTDETR are utilized by BoT-SORT tracking, which incorporates bridge-specific camera motion estimation and two-stage identity association. Precise vehicle positioning is achieved through dual-bounding-box localization, in which body-suspension error minimization and orientation vector updating are implemented. Experimentally, LC-RTDETR outperforms RTDETR with a 9.8% higher frames per second, 48.2% fewer parameters, and 65.4% lower floating-point operations. Practical validation confirms robustness to illumination changes, occlusion, motion blur, and adverse weather while accurately capturing stable trajectory during lane-changing maneuvers and speed fluctuations to enable vehicle localization. Finally, effective weight-position matching is fully integrated within the framework.