Improved Visual Odometer Based on Two-stream Convolutional Neural Networks

The purpose of the visual odometer is to use the image information input by the camera to estimate the motion information of the camera. As a positioning method, it has been widely used in fields such as autonomous driving and mobile robots in recent years. Because of the disadvantages of traditional visual odometers such as cumbersome implementation process and complicated calculation, the improvement of visual odometers has always been a hot research topic. In this paper, a two-stream convolutional neural networks is used, while RGB images and depth images are input to the model for training, and the Inception module is used to improve the convolutional layer and reduce the number of parameters. At the same time, an attention model is added to the convolutional layer to improve the network’s recognition of image features and the robustness of the system. Finally, the model was trained and tested on the KITTI data set. The results show that the model we proposed has achieved a certain improvement compared with VISO2-M and SfMLearner.