Sidewalk Extraction on Aerial Images With Deep Learning and Path Planning Algorithm

High-definition (HD) maps play an important role in autonomous driving by providing vehicles with localization functionality, environmental information, and support its mission and motion planning. In an HD map, road network extraction/annotation is a crucial feature that helps the autonomous vehicle navigate and keeps it within the safe driving zone. While road network extraction traditionally focuses on motorways and their boundaries, the emergence of small-scale autonomous vehicles, such as delivery and service robots, has shifted attention to sidewalks. Sidewalks are critical for safe and efficient navigation in residential and urban areas, yet automated methods for sidewalk extraction remain underexplored. To address this gap, this article proposes a sidewalk extraction method on aerial images using deep learning with the transfer learning technique. A path-planning algorithm-based refinement method is also proposed to further refine the extracted sidewalk. The proposed method can precisely extract sidewalks from aerial photographs and fix sidewalk discontinuity issues caused by occlusions. A sidewalk dataset is also explicitly designed for sidewalk extraction and semantic segmentation research. This article’s work is meant to fill the sidewalk extraction gap for road network extraction.