Improving Long Term Accuracy of Visual Localization in Urban Environment

Abstract

Localization remains a pivotal aspect of mobile robotics, with robots being required to discern their position by comparing sensor inputs against a pre-established environmental map. Notably, environmental shifts over time can diminish the reliability of these localization efforts. To address this challenge, our study introduces two interventions: dynamic object masking and CNN model fine-tuning, both scrutinized through extensive real-world experiments involving a robot operating continuously over a four-month span. Such long-duration testing of mobile robot localization in fluctuating environments is scarcely reported in the existing literature, where most focus on localization system usage on one shot or for a short period of time. Additionally, the robustness of localization can only be observed in real-world usage in the long term. Our findings reveal that while fine-tuning the CNN model may not drastically enhance the accuracy of the immediate testing environment, it significantly bolsters the robustness and accuracy of the system, especially when adapted to robots equipped with various sensor technologies. Our experiment showed improvement in various metrics, with the most noticeable improvement from 84.5% to 87.5% in positional accuracy. The novelty of this work is that it provides rare empirical evidence of extended operational challenges and solutions and unveils a deeper understanding pivotal for advancing mobile robot localization under the inevitable condition of environmental change.

Keywords

• Mobile robots
• robot localization
• visual place recognition