Dan Qiao, Liang Chen, Sheng Jin, and Yueyuan Zhang
Simultaneous localisation and mapping (SLAM), Loop closure detection (LCD), domain-invariant module, image translation model.
In visual simultaneous localisation and mapping (SLAM) methods, loop closure detection (LCD) is used to detect whether a scene has been visited or not, which can reduce the uncertainty of the estimated pose and map. However, a main challenge in LCD is the domain shift caused by environmental changes in day and night lighting, different weathers and seasons. To reduce the domain gap, this paper proposes a novel cross-domain method for LCD based on an image translation model CycleGAN+. Specifically, a domain-invariant module in CycleGAN+ is used to produce strongly discriminative domain-invariant images from input images of different domains by enforcing novel constraints in the loss function. Experimental results in five public datasets with challenging environmental changes show that the proposed method can improve the performance of cross-domain LCD tasks. The code is available at: https://github.com/Qd66666/Improve-CycleGAN.
Important Links:
Go Back
