|Synthetic Deep Neural Network Design for Lidar-inertial Odometry Based on CNN and LSTM
Hyunjin Son, Byungjin Lee, and Sangkyung Sung*
International Journal of Control, Automation, and Systems, vol. 19, no. 8, pp.2859-2868, 2021
Abstract : This paper proposes an integrated navigation algorithm based on the deep learning method using lidar and inertial measurements. The proposed method develops a new synthetic structure of neural networks for implementing the Lidar-inertial odometry to generate a 6 degree of freedom pose estimation. The proposed network consists of component neural networks that reflect each sensor’s characteristics, then an integrating network for combining estimates from heterogeneous sensors at the terminal stage. To secure an efficient estimation performance, a compound loss function design is exploited. The performance of the proposed deep learning-based LIO algorithm was verified through artificially generated data sets based on a high fidelity dynamics simulator. Instead of using the well-known reference data set of ground vehicles, the employed data set reflects the full 3D dynamic characteristics of the drone as well as low-cost sensor characteristics considering onboard implementation. Through the flight simulator data set, the estimation performance of the proposed synthetic network was demonstrated.
Deep learning, Lidar-inertial odometry, loss function, pose estimation, synthetic neural network.