DAVO:A Monocular Visual Odometry Method Based on Dual Attention

Abstract

In recent years, Visual odometry(VO) has been widely used in fields such as autonomous driving and virtual reality. Traditional methods for solving visual odometry rely on complex processes such as feature extraction, feature matching and camera calibration, and have low robustness and serious accuracy deficiency problems in challenging environments. In this paper, we propose a dual attention monocular visual odometry model that integrates Deep Learning(DL) with Reinforcement Learning(RL), named DAVO (Dual Attention Visual Odometry). The model combines a recurrent attention network model with a self-attentive mechanism to solve the relative poses of six degrees of freedom(6-DoF) by learning the image region locations that are favorable for the model pose estimation through a reinforcement learning algorithm. Finally, the model is evaluated and compared on the publicly available dataset KITTI. Compared with other mainstream models, DAVO only inputs 14.04% of the data in the image preprocessing stage, runs faster and outperforms most of the mainstream models.