Enhancing Visual Odometry: Integrating Load Variability and Optimized RGB-Depth Fusion in Transformer Networks

Visual Odometry provides a robust alternative to unreliable GPS and compass sensors, particularly for indoor navigation. We propose an improved Visual Odometry Transformer framework tailored for autonomous agents operating under diverse load conditions. Load variations significantly affect the agent's motion dynamics, leading to inaccuracies in turn angles and forward distances with traditional methods. To address this, we introduce a load variable during training, enabling the model to account for these variations. Furthermore, we optimize the input structure to reduce the training time and expand the dataset with random load conditions, achieving a navigation success rate of 89.4%, surpassing previous methods. These advancements highlight the potential of visual odometry to maintain robust performance in complex environments with varying loads, without increasing computational costs.