A review of computer vision-based algorithms for vehicle load spatial-temporal distribution identification

Vehicle loads play a critical role in the performance and safety assessment of in-service bridges, making accurate perception of their spatial-temporal distribution a fundamental task in bridge health monitoring. However, traditional contact-based and indirect measurement approaches are constrained by high costs, limited spatial coverage, and insufficient real-time capability. In contrast, computer vision technologies provide a promising non-contact alternative for vehicle load perception. This paper presents a comprehensive review of computer vision-based algorithms for identifying vehicle load spatial-temporal distributions on bridges, tracing the technological evolution from basic vehicle detection and tracking to advanced load modeling and post-processing techniques. First, traditional vehicle weighing methods are contrasted with computer vision-assisted bridge weigh-in-motion systems, highlighting the paradigm shift from contact-based measurements to visual perception. Subsequently, mainstream vehicle detection and tracking algorithms are reviewed and compared, and their applicability to different traffic monitoring scenarios is discussed. The core of the review focuses on three hierarchical levels of vehicle load modeling and perception, including equivalent point loads, axle loads, and wheel loads, the corresponding identification algorithms are examined in detail, demonstrating increasing levels of refinement achieved through detection, segmentation, and geometric inference. Furthermore, key post-processing challenges in vehicle load spatial-temporal distribution identification are analyzed in a structured manner, including homography estimation, trajectory optimization, and robustness under complex conditions such as lighting variation and occlusion. Through a comparative analysis of existing approaches, the advantages of deep learning-based methods over traditional digital image processing techniques are highlighted, while current limitations, particularly in terms of computational efficiency and generalization to the specific load distribution identification tasks, are also identified. Finally, future research directions are proposed, emphasising the development of efficient model architectures, benchmark datasets, and algorithmic adaptability. This review aims to serve as a reference for researchers and engineers engaged in the development and application of vision-based vehicle load spatial-temporal identification systems. © 2026 IOP Publishing Ltd. All rights, including for text and data mining, AI training, and similar technologies, are reserved.