Global Descriptor from Ceiling-Pointed Fisheye Image for Visual Place Recognition

This paper presents a visual place recognition (VPR) approach for indoor mobile robot localization using a ceiling-pointed fisheye camera. Unlike conventional forward-facing cameras, upward-facing fisheye cameras provide a wide, occlusion-free view of the upper hemisphere, which is typically more stable and less dynamic. However, fisheye distortion and the lack of annotated datasets pose significant challenges to existing global descriptor methods. To address these issues, we reframe VPR as an image retrieval problem and fine-tune the NetVLAD architecture to learn robust global descriptors from distorted ceiling-pointed fisheye images. To overcome dataset scarcity, we incorporate a synthetic dataset generated using NVIDIA Isaac Simulator, covering diverse indoor environments and lighting conditions. Our approach is evaluated on both synthetic and real-world datasets. Experimental results show that the trained Fisheye NetVLAD consistently outperforms the baseline NetVLAD, achieving higher Recall@K values across all ranks. These results highlight the potential of using a single fisheye camera with learned descriptors for scalable, cost-effective, and efficient indoor robot localization.