Abstract
This study introduces a novel Stereo Attentionbased Visual Inertial Odometry model, namely SATVIO, aiming to enhance odometry performance by leveraging deep learning techniques for sensor fusion. The research evaluates the SATVIO model against existing Visual Odometry methods using the KITTI odometry dataset, focusing on translational and rotational accuracy enhancements through innovative attention mechanisms and early fusion strategies. The proposed model integrates Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks to process and fuse data from stereo image inputs and inertial measurements effectively. SATVIO model particularly employs Triplet Attention (TA) and early fusion techniques with the aim of addressing the challenges posed by scale ambiguity and environmental changes. The results demonstrates that our proposed model outperforms traditional methods in specific configurations, thus demonstrating competitive or superior performance on key challenging sequences.
Funding Information
  • Department of Computer Science and Media Technology
  • Sustainable Digitalization Research Center at Malmö University