SDF-Guided Multi-modal Big Data Road Extraction

Kun Feng, Juanjuan Zhao, Kejiang Ye, Xitong Gao

Abstract: Automatic extraction of road networks from multimodal big data sources has been a longstanding challenge. GPS data and satellite imagery are commonly used data sources for road extraction tasks due to their widespread availability and ease of access. However, existing methods often rely on binary segmentation labels, and the discontinuity inherent in binary segmentation increases the difficulty of predicting road boundaries accurately. In contrast, we introduce the use of the Signed Distance Field (SDF), which can transform the discontinuous binary 0/1 pixel labels into continuous pixel values ranging from −1 to 1. This approach allows the SDF labels to complement the binary labels in guiding the model during the segmentation task. Subsequently, we compute regression and segmentation losses for the model predictions with respect to both types of labels, and integrate these losses for backpropagation. We evaluated our method on popular dataset from Shenzhen, achieving state-of-the-art results. During our experiments, we discovered that GPS data not only exhibits common positional noise but also temporal noise, manifesting as multiple GPS points sharing the same timestamp. Temporal noise in GPS data poses significant challenges to algorithms that rely on temporal information, such as KDE. We analyzed the temporal noise present and proposed a temporal noise filtering algorithm. Our algorithm significantly improves the performance of previous algorithms that rely on temporal consistency.

External IDs:doi:10.1007/978-981-96-8298-0_10