A monocular-based framework for accurate identification of spatial-temporal distribution of vehicle wheel loads under occlusion scenarios

Ensuring the safety and durability of roads and bridges necessitates the accurate identification of spatial-temporal distribution of vehicle loads. This can be achieved using computer vision algorithms, but these require vehicles to be unoccluded, a condition often unmet in dense traffic. This study introduces an innovative framework designed for the precise determination of the spatial-temporal distribution of vehicle wheel loads, particularly tailored to accommodate scenarios involving occlusion. The initial phase involves the application of an object detection model and a keypoint detection model to identify vehicles within surveillance footage and ascertain the precise locations of wheel load points. Thus, a wheel load position monitoring algorithm is developed specifically for unoccluded scenarios. Subsequently, to address situations where the target vehicle is occluded, an image binary classification model is employed. Employing class activation map technology, an annotation-free occlusion region mask generation algorithm is proposed to cover occluded areas. The key contribution lies in the integration of an image inpainting model, enabling the transformation of occluded vehicle images into unoccluded ones through a mask-inpainting operation. This strategy achieves accurate identification of vehicle wheel load positions irrespective of occlusion. A field test conducted to assess the performance of this algorithm revealed a substantial improvement in identification accuracy of occluded vehicle wheel loads post image mask-inpainting model processing, achieving accuracy comparable to unoccluded cases.