Abstract: Our deep visual Odometry (VO) approach prioritizes memory and improved postures while taking into account global information. Unlike existing learning-based approaches which treat VO as a simple tracking issue, we reconstruct camera postures from picture fragments, which leads to a high accumulation of errors. To correct past mistakes, accurate worldwide data is essential. However, it might be difficult for end-to-end systems to reliably store such data. Therefore, we have developed an adaptive memory module that preserves information gradually and adaptively from a local to a global level in a neural memory analogue. This approach has been further enhanced using a refining module that takes advantage of global information stored in memory. To pick features for each view based on the co-visibility in the feature domain, we use past outputs as a guide and apply a spatial-temporal attention. Our system is more advanced than simple tracking since it has a memory module and a refinement module. Our experiments on the KITTI and TUM-RGBD datasets demonstrate that our technique not only delivers competitive results compared to conventional approaches in normal settings but also significantly outperforms state-of-the-art methods. Moreover, our model performs very well in instances where traditional algorithms struggle, such as texture-less areas and sudden movements.

| DOI: 10.17148/IARJSET.2023.10835