VIOFormer: Advancing Monocular Visual-Inertial Odometry Through Transformer-Based Fusion

Visual-Inertial Odometry (VIO) algorithms are pivotal for localization and navigation, with deep learning enhancing their potential. However, long-distance pose estimation still poses challenges. While most VIO algorithms use Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), the promise of Transformer models remains largely unexplored. We present VIOFormer, a new monocular visual-inertial pose estimation technique harnessing Transformers. By leveraging their attention mechanism and long-range dependencies, VIOFormer addresses the cumulative error issue linked with IMU and augments pose accuracy. Its architecture combines the preprocessing of adjacent frame sequences with IMU data, an encoder, and a decoder, allowing it to discern features from both visual and IMU data for precise pose estimation. Testing on the KITTI dataset confirms VIOFormer's superiority over both traditional and contemporary deep learning-based VIO approaches.