Author(s) :

Shamsiddinov Nodir.

Volume/Issue :

Abstract :

Visual odometry systems, technologies that help machines understand their position by analysing camera images, are essential for safety-critical applications like autonomous vehicles and robotic navigation. These systems work well in controlled laboratory conditions, but they often struggle in realworld deployment, where cameras face challenges such as sensor readout noise and motion blur. The key problem is that training data used to teach these systems is often too clean and simplified, leaving them unprepared for challenging real-world conditions. This study demonstrates that the systematic injection of realistic image corruptions during training reduces median translation error by 20–70% relative to a non-augmented baseline, substantially improving model robustness across a range of degradation scenarios. Five types of synthetic noise were evaluated—Gaussian noise, Poisson noise, motion blur, spatially correlated noise, and combinations of these—comparing models trained on clean data against those trained with augmented data, all using an optimization approach called Conservative Pose Loss. Noise-augmented models consistently achieved lower position and orientation errors than the baseline, and this advantage was maintained even when encountering corruption types not seen during training. These results demonstrate that incorporating realistic noise patterns into training data is crucial for deploying reliable visual odometry systems in safety-critical real-world applications. These findings suggest that careful attention to data augmentation strategies can bridge the gap between laboratory performance and practical deployment, potentially improving the safety and reliability of autonomous systems navigating complex environments.