Enhancing Visual Inertial Odometry Performance using Deep Learning-based Sensor Fusion

Odometry estimation plays a key role in facilitating autonomous navigation systems. While significant consideration has been devoted to research on monocular odometry estimation, sensor fusion techniques for Stereo Visual Odometry (SVO) have been relatively neglected due to their demanding computational requirements, posing practical challenges. However, recent advancements in hardware, particularly the integration of CPUs with dedicated artificial intelligence units, have alleviated these concerns. In this paper, we investigate the efficacy of attention mechanisms and the incorporation of stereo input in comparison to monocular odometry, aiming to enhance the performance of SVO. We tested two different types of attention mechanisms, i.e., Triplet Attention (TA) and Convolutional Block Attention Module (CBAM), and their fusion in two stages Early and Late. Our results show that the fusion of the second camera improves the performance of the model, as well as early fusion with TA provided the best results.