Enhancement of Building Change Detection based on ChangeFormer through Integration of Deformable Attention Transformer

Keywords: change detection, transformer, deformable attention
Abstract Type: Paper Abstract

Abstract

The evolution of autonomous driving systems, coupled with an increasing dependence on high-definition mapping, has amplified the need for advanced technologies to efficiently detect alterations in urban structures and terrain, enabling timely map updates. This study introduces a model that enhances the accuracy of building change detection using satellite imagery by integrating a Deformable Attention Transformer (DAT) module into the ChangeFormer framework. The proposed model significantly reduces approximately 14 million parameters compared to the original ChangeFormer while simultaneously improving performance metrics. The synergy between the ChangeFormer architecture and the DAT module effectively addresses challenges in change detection with greater efficiency and precision than conventional models that rely solely on convolutional networks and standard transformer components. These findings are promising for advancing precise map update technologies, particularly in building change detection, thereby supporting the development of dynamic, high-fidelity mapping solutions for autonomous navigation systems.