Abstract
This paper presents an adaptive sensor fusion framework for Global Positioning System (GPS)-denied navigation that relies on monocular camera and Inertial Measurement Unit (IMU) data. Unlike classical tightly coupled visual–inertial odometry (VIO) pipelines, the proposed architecture computes independent pose estimates from camera and IMU streams and employs a lightweight neural reliability predictor to combine them. During a GPS ground truth is available, pose errors of both sensors are computed and fed to a compact multi-layer perceptron. Using a Gaussian negative log-likelihood loss, the network learns the future error distribution of each sensor in terms of mean and variance. In the deployment phase, after GPS becomes unavailable, these predicted statistics are converted into dynamic reliability weights via an inverse-variance scheme, yielding a per-frame estimate of how much each sensor should influence the fused pose. Experiments on the KITTI dataset indicate that this strategy behaves as intended: when IMU-based pose exhibits larger rotational errors, visual odometry compensates for orientation drift, whereas in segments where translational accuracy of the IMU is higher, the fused trajectory leans toward the inertial solution. Overall, the results demonstrate that modeling time-varying sensor reliability can provide more consistent pose estimates than relying on a single sensor modality.