A High-Precision, Low-Latency, and Robust Vision-Based Localization System for Embodied Intelligent Robots in Complex Environments

Vision-based localization is essential for embodied intelligent robots, providing real-time position and orientation estimates with high precision and low cost. However, existing approaches struggle in complex environments and on resource-constrained embodied robots due to unreliable feature correspondences between captured images and localization databases. To address this challenge, this demo presents a high-precision, low-latency, and robust vision-based localization system designed for embodied intelligent robots operating in complex environments. The system mitigates unreliable feature correspondences by integrating a series of methods and benchmarks for detecting, describing, matching, filtering, and evaluating feature points and line segments. These methods are computationally efficient and effective in complex environments, while the benchmarks provide a reliable evaluation framework for vision-based localization. Experiments using natural datasets demonstrate that the system achieves centimeter-level positioning accuracy and sub-degree orientation accuracy, even in challenging outdoor environments. Supporting materials for this demo are available at https://drive.google.com/drive/folders/1udyeiyw1l7SCP70v2pXEL2klKHnQY43M?usp=sharing.