IRPL-VO: illumination-robust visual odometry based on multi-stage point-line feature interaction

Abstract Visual odometry (VO) has been widely applied in autonomous driving, robotic navigation, and unmanned aerial vehicle localization, where accuracy and robustness are critical to overall system performance. However, most existing VO systems struggle in illumination-challenging scenarios, including overexposure and low-light scenarios. In these scenarios, extracted visual features often become sparse and unstable, resulting in unreliable feature matching and frequent tracking failures. To address these challenges, we design a VO system that simultaneously extracts point and line features, with multi-stage interactions designed between them. Specifically, the multi-stage feature interaction is introduced to fully exploit the potential correlations between point and line features. Furthermore, we introduce the line feature enhancement module, which leverages parallel standard and difference convolution layers to extract richer and more complete line features, thereby improving robustness to illumination variations. Extensive experiments on both public and self-constructed datasets demonstrate that the proposed VO system consistently outperforms other advanced VO methods across diverse illumination-challenging environments.