TSFNet: A Two-Stage Fusion Network for Visual–Inertial Odometry

In autonomous operations of unmanned aerial vehicles (UAVs), accurate pose estimation is a core prerequisite for achieving autonomous navigation, obstacle avoidance, and task execution. To address the challenge of localization in GNSS-denied environments, Visual–Inertial Odometry (VIO) has emerged as a mainstream solution due to its outstanding performance. However, existing deep learning-based VIO methods exhibit limitations in their multi-modal fusion mechanisms. These methods typically employ simple concatenation or attention mechanisms for feature fusion. Furthermore, enhancements in accuracy are often accompanied by significant computational overhead. This makes it difficult for models to effectively handle complex, dynamic scenes while remaining lightweight. To this end, this paper proposes TSFNet (Two-stage Sequential Fusion Network), an efficient two-stage sequential fusion network. In the first stage, the network employs a lightweight visual backbone and a bidirectional recurrent network in parallel to extract spatial and motion features, respectively. A gated fusion unit is employed to achieve adaptive intra-frame feature fusion, dynamically balancing the contributions of different modalities. In the second stage, the fused features are organized into sequences and fed into a dedicated temporal network to explicitly model inter-frame motion dynamics. This decoupled fusion architecture significantly enhances the model’s representational capacity. Experimental results demonstrate that TSFNet achieves superior performance on both the EuRoC and Zurich Urban MAV datasets. Notably, on the Zurich Urban MAV dataset, it reduces the localization Root Mean Square Error (RMSE) by 62% compared to the baseline model, while simultaneously reducing the number of parameters and computational load by 76.65% and 24.30%, respectively. This research confirms that the decoupled two-stage fusion strategy is an effective approach for realizing high-precision, lightweight VIO systems.