When training samples of existing deep learning-based Visual Odometry (VO) are different from application scenarios, it is difficult to adapt to the new environment. Therefore, an online updated monocular visual mileage calculation method(OUMVO) is proposed in this paper. In the application stage, the pose estimation network model is optimized online by using the real-time image sequence, which improves the generalization ability of the network and the ability to apply to the new environment. At the same time, self-supervised learning method is utilized without the need to mark the ground truth. Moreover, Transformer is adopted to conduct sequential modeling of image streams to make full use of the visual information within the local window to improve the precision of the pose estimation in order to avoid the limitation that the traditional method can only use two adjacent frames to estimate the pose. It can also compensate for the shortcomings of using RNN for sequence modeling which can not be calculated in parallel. In addition, it uses the geometric consistency constraint of the image space to solve the scale drift problem of the traditional monocular visual mileage calculation method. Quantitative and qualitative experimental results on the KITTI dataset show that the proposed method is superior to existing state-of-the-art monocular visual odometry methods in terms of pose estimation accuracy and adaptability to new environments.

[2023-12-29] Accepted Paper /doc/2023.09.0512/v1