MDAG-Net: Multidomain Association-Guided Network for Image-Based Long-Term Visual Localization

In the case of long-term changing environment, long-term visual localization is a challenging problem in autonomous driving and mobile robots. Due to the influence of season, illumination and other changing weather conditions, the traditional image retrieval methods are difficult to achieve ideal results in long-term visual localization. Therefore, inspired by the human brain associative recognition function, an image retrieval based on a multidomain association-guided network is proposed to solve the long-term visual localization problem. The key idea is to extract the discriminative domain-invariant features in different scenes through multidomain image transformation of the perceptual network and the conceptual network. In addition, in order to better associate image features of different scenes in the conceptual network and guide the perceptual network to obtain more robust domain invariant features, an association-guided module is designed without the need for external datasets. On this basis, the domain feature loss function and the guidance mechanism of the loss function are introduced to assist these two network models training to obtain better performance. Finally, experiments are carried out on the CMU-Seasons dataset and the RobotCar-Seasons dataset. Compared with some state-of-the-art methods, the proposed method improved the high-precision localization result of urban, suburban, and park scenes in the CMU-Seasons dataset by 1.5%, 0.5%, and 0.7%, respectively, which also can verify the effectiveness of the proposed method under various seasonal and illumination conditions.