GAN-SLAM: GAN based Monocular Visual-Inertial Simultaneous Localization and Mapping in Dark Environments

In recent years, visual-inertial Simultaneous Localization and Mapping (SLAM) techniques have achieved excellent results due to the complementary nature of visual and inertial sensors, the challenges remain in some environments. When encountering a challenging environment, image features become an important factor affecting the success rate and accuracy of localization of SLAM system. Due to the Convolutional Neural Network (CNN) has the advantage of extracting effective features more accurately, this paper makes a new attempt to replace traditional features with those generated by the network. We combine the state-of-the-art CNN of optical flow estimation and iterative optimization model of traditional SLAM framework, which achieves the accuracy and robustness in challenging environments. To show the performance of our method, some state-of-the-art visual-inertial methods are tested and compared with the challenging sequences of public dataset. Experimental results show that the proposed method can achieve better results in accuracy and robustness and works reliably well for challenging environments.