Learning-based IMU bias prior for robust visual-inertial odometry in degraded environments

Visual-Inertial Odometry (VIO) has become a fundamental component in autonomous navigation for mobile robots, offering low-cost and high-accuracy localization by fusing camera and IMU measurements. However, most existing VIO systems are highly dependent on visual observations and often degrade significantly in feature-deficient or visually challenging environments, such as foggy, smoky, low-light, or motion-blurred visual conditions. This work proposes a plug-and-play bias prior method to enhance the robustness and accuracy of VIO under degraded visual conditions by leveraging data-driven IMU bias estimation. We first construct a supervised training dataset based on the EuRoC dataset, where IMU bias labels are derived from ground-truth poses via pre-integration and subsequently smoothed to serve as supervision. A lightweight temporal network is then designed to regress the 6-DoF IMU bias and its uncertainty from raw IMU sequences. The network architecture integrates residual 1D convolutions, multi-head self-attention, and stacked GRU modules to extract both local temporal features and global dynamics. The predicted mean and uncertainty are further encoded as a residual prior factor and incorporated into the VINS-Mono backend optimization in a weighted least-squares form. The influence of this factor is jointly determined by the learnable uncertainty and manually adjustable global weights, thereby achieving a balance between visual and inertial constraints. To address the challenges posed by real-world environments, this dissertation introduces periodic Gaussian blur, illumination changes, and texture dropout to construct a degraded version of the EuRoC dataset, simulating extreme visual conditions. Extensive experiments conducted on both the standard and degraded datasets demonstrate that the proposed method significantly improves localization accuracy and system robustness, particularly in scenarios where visual tracking performance degrades. Additionally, further analysis of weight configurations provides valuable insights for tuning the prior factor. The proposed method achieves effective generalization across multiple data sequences without altering the core structure of VIO, and through modular ROS integration, it supports application across various types of mobile platforms.