InertialNet: Toward Robust SLAM via Visual Inertial Measurement

SLAM (simultaneous localization and mapping) is commonly considered as a crucial component to achieving autonomous robot navigation. Currently, most of the existing visual 3D SLAM systems are still not robust enough. Image blur, variation of illumination, and low-texture scenes may lead to registration failures. To let the visual odometry (VO) deal with these problems, the workflow of traditional approaches becomes bulky and complicated. On the other hand, the advancement of deep learning brings new opportunities. In this paper, we use a deep network model to predict complex camera motion. It is different from previous supervised learning VO researches and requires no camera trajectories which are difficult to obtain. Using the image input and IMU output as an end-to-end training pair makes data collection cost-effective. The optical flow structure also makes the system independent of the appearance of training sets. The experimental results show that the proposed architecture has faster training convergence, and the model parameters are also significantly reduced. Our method is able to remain certain robustness under image blur, illumination changes, and low-texture scenes. It can correctly predict the new EuRoC dataset, which is more challenging than the KITTI dataset.