Enhancing VIO Robustness Under Sudden Lighting Variation: A Learning-Based IMU Dead-Reckoning for UAV Localization

Visual Inertial Odometry (VIO) is commonly used for real-time Unmanned Aerial Vehicle (UAV) localization. However, the performance of VIO significantly deteriorates when UAV encounters sudden lighting variation in the environment, which poses a significant risk during flight. To address this issue without introducing additional sensors, a learning-based dead-reckoning algorithm relying solely on inertial measurement, which shares the same source with VIO, is proposed. The core idea of our method tightly couples a model-based Left Invariant Extended Kalman Filter (LIEKF) with a statistical neural network, both driven by raw inertial measurement. We have validated our algorithm for comparable accuracy with commonly deployed VIO methods under favorable lighting conditions and outperforms other IMU dead-reckoning algorithms in open-source datasets and real-world scenarios. To further enhance localization robustness while UAV traverses environments with different lighting conditions, we introduce an approach that tightly integrates our algorithm with VIO, and validate its effectiveness in real-world scenarios. It is believed that our work presents a promising way for enhancing robustness in vision-based localization methods within the robotics society.