Spatial Gating with Hybrid Receptive Field for Robot Visual Localization

Abstract

Abstract Visual localization for mobile robots is a sophisticated and challenging task that necessitates the extraction of pertinent scene information from images obtained by the robot’s visual sensors to ascertain its position within an environment. The task is complicated by variations in environmental factors which affect the accuracy of localization. To address the challenges of visual localization on variations of illumination, seasons, and viewpoints, this paper proposes a visual localization network based on a gated selection and hybrid receptive field. We utilize a fine-tuned DINOv2 for local feature extraction and leverage a hybrid receptive field to enhance the diversity of visual features. Furthermore, our approach employs spatial gating to dynamically and effectively select and aggregate the advantageous spatial features. Extensive experiments demonstrate that the visual localization performance of our approach surpasses existing methods on multiple challenging datasets, particularly achieving a Recall@1 metric of 69.2% on the NordLand dataset, which signifies a 10.8% enhancement compared to MixVPR.

Keywords

• Mobile robot
• Visual localization
• Gating network
• Global descriptor
• Image retrieval