Monocular VO Based on Deep Siamese Convolutional Neural Network

Abstract

Deep learning-based visual odometry systems have shown promising performance compared with geometric-based visual odometry systems. In this paper, we propose a new framework of deep neural network, named Deep Siamese convolutional neural network (DSCNN), and design a DL-based monocular VO relying on DSCNN. The proposed DSCNN-VO not only considers positive order information of image sequence but also focuses on the reverse order information. It employs supervised data-driven training without relying on any modules in traditional visual odometry algorithm to make the DSCNN to learn the geometry information between consecutive images and estimate a six-DoF pose and recover trajectory using a monocular camera. After the DSCNN is trained, the output of DSCNN-VO is a relative pose. Then, trajectory is recovered by translating the relative pose to the absolute pose. Finally, compared with other DL-based VO systems, we demonstrate the proposed DSCNN-VO achieve a more accurate performance in terms of pose estimation and trajectory recovering through experiments. Meanwhile, we discuss the loss function of DSCNN and find a best scale factor to balance the translation error and rotation error.