Investigation on change detection of high-definition (HD) map based on sequential model

Due to the precise localisation capabilities and rich semantic features, high-definition (HD) maps have gained significant attention in the fields such as autonomous vehicle and unmanned aircraft. However, their high temporal sensitivity leads to substantial maintenance costs, making efficient and accurate map updates a major challenge. HD map change detection enables rapid identification of altered regions, significantly improving update efficiency and reducing manual verification costs, thus facilitating map maintenance. This paper proposes a sequential neural network framework for HD map change detection (SeqHDMapCDNet), in order to provide real-time HD maps for autonomous vehicles and enable rapid updates. The framework leverages Differential Feature Fusion module (Diff Module) to learn multi-scale differential features by comparing changes between the old and new maps. Subsequently, we integrate these multi-scale differential features and achieve temporal fusion through Convolutional Long Short-Term Memory (ConvLSTM). Finally, multi-scale inference is employed to expedite the model convergence. On our newly created dataset, SeqHDMapCDNet has shown significant improvements. Specifically, intersection over Union (mIoU), mean F₁ score (mF₁), and overall accuracy (OA) have increased by 10.22%, 7.79% and 0.118%, respectively, surpassing several methods referenced in the literature. This optimisation demonstrates the superior performance of our model in the context of map updates.