A Cross-Spectral Image Fusion Based Visual Odometry

This paper proposes a cross-spectral image fusion based visual odometry with GAN (Generative Adversarial Networks) and Residual Dense Blocks. It can well preserve the intensity information in infrared images and the gradient information in visible images. Based on fused images, the pose can be obtained. Experimental verification was performed on All-Day Visual Place Recognition Benchmark Dataset. And results in experiment show that compared to merely infrared and visible images, the positioning accuracy of the fusion image has been greatly improved.