Real-Time Spatial Importance Estimation Using Window-Based Spatiotemporal Aggregation

Urban digital twins support artificial intelligence (AI)-driven urban services by utilizing city-scale point cloud. However, real-time acquisition and processing of point clouds involve substantial communication and computational overhead. A key strategy to reduce these costs is to estimate salient regions as spatial importance within real-time point cloud frames. This article proposes real-time spatial importance estimation based on spatiotemporal changes. The proposed method integrates three components: spatial feature extraction using both static and dynamic approaches to enable informative extraction, centroid-based importance estimation that enables real-time processing, and window-based spatiotemporal aggregation to enhance the reliability of the estimation. Our experiments evaluate the effectiveness of the proposed method across various environments and devices. The results demonstrate that the centroid-based importance estimation consistently achieves real-time performance and mitigates device-dependent differences by utilizing a spatiotemporal window. Furthermore, in an object detection scenario, the proposed method significantly reduces the number of points in each frame while preserving detection accuracy. Overall, the proposed method shows strong potential for accurate and efficient real-time spatial importance estimation in AI-driven services.