| Summary: | 碩士 === 國立中正大學 === 資訊工程研究所 === 106 === To achieve autonomous robot navigation, SLAM (simultaneous localization and mapping) is a crucial component. For more than two decades of research, 2D-SLAM has already been deployed to the industry. Nowadays, we want the robot to have even more complete perception abilities, therefore more and more researchers put their effort on 3D dense SLAM.
However, most of the existing visual 3D SLAM systems are not robust enough. Image blur, variation of illumination, and low-texture scenes may lead to registration failures.
In order to deal with these problems, the work flow of traditional approaches become bulky and complex. On the other hand, the advancement of deep learning brings new opportunities.
We use a deep network model to predict complex camera motion, which is different from previous supervised learning VO researches, and requires no camera trajectories that are difficult to obtain.
Using image input and IMU output as end-to-end training pair makes data collection cost-effective. The optical flow structure also makes the system not depend on the appearance of training sets.
The experimental results show that the proposed architecture has a faster training convergence than the similar research, and the model parameters are also greatly reduced. It can also correctly predict the EuRoC dataset that is more challenging than KITTI dataset.
Our method could remain certain robustness under image blur, illumination changes and low-texture scenes.
Based on this result, future research can integrate prediction with the VIO system to construct a more robust visual SLAM system.
|
|---|
