Pixel to Elevation: Learning to Predict Elevation Maps at Long Range using Images for Autonomous Offroad Navigation

• DOI: 10.48550/arxiv.2401.17484
• Publication Date: 2024-01-30
• Authors: Chanyoung Chung; Georgios Georgakis; Patrick Spieler; Curtis Padgett; Shehryar Khattak
• Keywords: Elevation (ballistics)

Understanding terrain topology at long-range is crucial for the success of off-road robotic missions, especially when navigating high-speeds. LiDAR sensors, which are currently heavily relied upon geometric mapping, provide sparse measurements mapping greater distances. To address this challenge, we present a novel learning-based approach capable predicting elevation maps using only onboard egocentric images in real-time. Our proposed method comprised three main elements. First, transformer-based encoder introduced that learns cross-view associations between views and prior bird-eye-view map predictions. Second, an orientation-aware positional encoding to incorporate 3D vehicle pose information over complex unstructured with multi-view visual image features. Lastly, history-augmented learn-able embedding achieve better temporal consistency predictions facilitate downstream navigational tasks. We experimentally validate applicability our autonomous offroad navigation real-world driving data. Furthermore, qualitatively quantitatively compared against current state-of-the-art methods. Extensive field experiments demonstrate surpasses baseline models accurately while effectively capturing overall long-ranges. Finally, ablation studies conducted highlight understand effect key components their suitability improve capabilities.