Optimizing Odometry Accuracy through Temporal Information in Self-Supervised Deep Networks

In this paper, we propose a self-supervised deep recurrent convolutional network with a self-attention pipeline for LiDAR-Inertial odometry estimation that leverages temporal information. In contrast to existing learning-based approaches, our network does not require expensive ground truth and it utilizes sliding window optimization for self-supervisory signal for training. The results indicate that the proposed network is able to achieve accuracy levels comparable to conventional model-based methods and outperform other learning-based approaches that do not incorporate temporal information. Our findings provide strong evidence for the potential of incorporating temporal information in deep networks for LiDAR-Inertial odometry and highlight the effectiveness of our proposed self-supervised approach.